Portland State University PDXScholar

Dissertations and Theses Dissertations and Theses

2009 Analysis and Classification of Sounds Produced by Asian Elephants (Elephas Maximus)

Sharon Stuart Glaeser Portland State University

Follow this and additional works at: https://pdxscholar.library.pdx.edu/open_access_etds

Part of the Animal Sciences Commons, and the Biology Commons Let us know how access to this document benefits ou.y

Recommended Citation Glaeser, Sharon Stuart, "Analysis and Classification of Sounds Produced by Asian Elephants (Elephas Maximus)" (2009). Dissertations and Theses. Paper 4066. https://doi.org/10.15760/etd.5951

This Thesis is brought to you for free and open access. It has been accepted for inclusion in Dissertations and Theses by an authorized administrator of PDXScholar. Please contact us if we can make this document more accessible: [email protected]. THESIS APPROVAL The abstract and thesis of Sharon Stuart Glaeser for the Master of Science in

Biology were presented May 5, 2009, and accepted by the thesis committee and the department.

COMMITTEE APPROVALS:

RandyZeli~

Luis Ruedas

<::;;> Thomas Dolan

David Shepherdson

DEPARTMENT APPROVAL:

Michael Murphy, Chair Department of Biology ABSTRACT

An abstract of the thesis of Sharon Stuart Glaeser for the Master of Science in

Biology presented May 5, 2009.

Title: Analysis and classification of sounds produced by

Asian elephants (Elephas maximus)

Relatively little is known about the vocal repertoire of Asian elephants (Elephas maximus), and a categorization of basic call types and modifications of these call types by quantitative acoustic parameters is needed to examine acoustic variability within and among call types, to examine individuality, to determine communicative function of calls via playback, to compare species and populations, and to develop rigorous call recognition algorithms for monitoring populations.

This study defines an acoustic repertoire of Asian elephants based on acoustic parameters, compares repertoire usage among groups and individuals, and validates structural distinction among call types through comparison of manual and automated classification methods. Recordings were made of captive elephants at the Oregon Zoo in Portland, OR, USA, and of domesticated elephants in Thailand.

Acoustic and behavioral data were collected in a variety of social contexts and environmental noise conditions. Calls were classified using perceptual aural cues plus visual inspection of spectrograms, then acoustic features were measured, then automated classification was run. The final repertoire was defined by six basic call types (Bark, Roar, Rumble, Bark, Squeal, Squeal, and Trumpet), five call combinations and modifications with these basic calls forming their constituent parts (Roar-Rumble, Squeal-Squeak, Squeak train, Squeak-Bark, and Trumpet­

Roar), and the Blow. Given the consistency of classifications results for calls from geographically and socially disparate subject groups, it seems possible that automated call detection algorithms could be developed for acoustic monitoring of

Asian elephants.

2 1ANALYSIS AND CLASSIFICATION OF SOUNDS PRODUCED BY

ASIAN ELEPHANTS (ELEPHAS MAX/MUS)

by

SHARON STUART QLAESER

A thesis submitted in partial fulfillment of the requirements for the degree of

MASTER of SCIENCE in BIOLOGY

Portland State University 2009 Dedication

This thesis is dedicated to my dear husband, Bobby, who offered his love and support throughout this adventure and encouraged me to share my heart with elephants. Acknowledgements First and foremost I would like to thank my thesis committee for their expertise and guidance: Dr. Randy Zelick, Dr. Luis Ruedas, Dr. Thomas Dolan,

Dr. David Shepherdson, and Dr. David Mellinger. I am especially thankful to my advisor, Dr. Zelick, who met the challenges of data collection with enthusiasm and was so patient with me. I am eternally grateful to the Oregon Zoo elephant handlers, without whom this study would not have been possible. I especially appreciate the support from those most involved, April Yoder, Bob Lee, Joe Sebastiani, Jeb Barsh, and Dimas Dominguez. Dr. Mellinger and Dr. Bolger Klinck, with the OSU/NOAA

Cooperative Institute for Marine Resource Studies, provided invaluable supervision and support with acoustic analysis at a critical time. I appreciate the support and encouragement from the Oregon Zoo Conservation Program staff, Anne W amer,

David Shepherdson, and Karen Lewis, and Deputy Director Mike Keele. I thank Dr.

Mandy Cook for insight and guidance with early analysis, and Jean Lea Hofheimer and Bobbi Estabrook for scoring calls. I thank the managers at the Royal Elephant

Kraal and Elephant Nature Park for the opportunity to record domesticated elephants in Thailand. Travel funding was granted by Portland State University Marie Brown and Academically-Controlled Auxiliary Activities Fund. I thank my loving family,

Meghan Martin, and other dear friends who embraced this project with me. Finally, I am eternally grateful to Nancy Scott and the late Dr. LEL "Bets" Rasmussen for welcoming me into the elephant research community and paving the road, and to the many elephant handlers, veterinarians, and elephants who have taught me so much.

ii TABLE OF CONTENTS

ACKNOWLEDGEMENTS ...... ii

LIST OF TABLES ...... v

LIST OF FIGURES ...... vi

CHAPTER I: INTRODUCTION ...... 1

TAXONOMY AND CONSERVATION STATUS ...... 2

ECOLOGY AND COMMUNICATION ...... 4

VOCAL PRODUCTION AND INDIVIDUALITY IN ACOUSTIC SIGNALS ...... 6

Low-FREQUENCY SOUNDS ...... 9

HEARING ...... 11

PURPOSE AND SIGNIFICANCE OF THIS THESIS RESEARCH ...... 12

CHAPTER II: PUTATIVE CALL TYPES AND MANUAL CALL CLASSIFICATION ...... 16

METHODS ...... 16

RESULTS ...... 21

CHAPTER III: CALL DISTRIBUTION AND CALL RATE - HOW THE REPERTOIRE IS USED BY GROUPS AND INDIVIDUALS ...... 27

METHODS ...... 27

RESULTS ...... 29

CHAPTER IV: ACOUSTIC PARAMETERS AND AUTOMATIC CLASSIFICATION OF CALLS ...... 33

METHODS ...... 33

RESULTS ...... 53

lll CHAPTER V: DISCUSSION ...... 78

REFERENCES ...... 82

APPENDIX A: DISTRIBUTION PLOTS OF PARAMETERS ...... 87

HISTOGRAMS AND BOXPLOTS OF ORIGINAL DATA ...... 87

Q-Q PLOTS OF SCALED DAT A ...... 104

Box PLOTS OF SCALED v ARIABLES FOR IDENTIFYING OUTLIERS ...... 111

APPENDIX B: CLASSIFICATION ON TREE RAW DATA OUTPUT ...... 118

APPENDIX C: PRINCIPAL COMPONENT ANALYSIS DATA - RAW DATA OUTPUT ...... 123

APPENDIX D: ANALYSIS OF SIMILARITY (ANOSIM) DATA OUTPUT ...... 124

APPENDIX E: POWER SPECTRA OF RUMBLES AND NOISE ...... 127

lV LIST OF TABLES

Table 1: Acoustic ethogram of putative call types showing single calls, trains of calls, and call combinations ...... 22

Table 2: Social contexts for call comparison ...... 28

Table 3: Acoustic parameters measured by Osprey ...... 38

Table 4: Calls in dataset for statistical analysis (N=927) ...... 42

Table 5: Acoustic parameters that were used in statistical analysis ...... 45

Table 6: Calls in dataset (N=908) after outliers removed ...... 49

Table 7: Confusion matrix for tree ...... 59

Table 8: Classification rate(%) for tree ...... 59

Table 9: Summary of pair-wise ANOS IM showing significant difference between all call types ...... 66

Table 10: Median values of acoustic parameters shown by the classification tree and PCA to differentiate basic call types ...... 67

Table 11: Final ethogram for Bark, Roar, Rumble, Squeak, Squeal, Trumpet, and Blow...... 68

v LIST OF FIGURES

Figure 1: Call distribution of the Oregon Zoo herd versus domesticated elephants in Thailand ...... 29

Figure 2: Call distribution of females in the Oregon Zoo herd ...... 30

Figure 3: Call distribution of one male (Tusko) in the Oregon Zoo herd ...... 31

Figure 4: Mean call rate as a function of social context ...... "...... 32

Figure 5: Osprey window of Squeak with annotation box and feature box ...... 34

Figure 6: Osprey window of Trumpet with annotation box and feature box ...... 35

Figure 7: Acoustic parameters measured by Osprey that can be visualized ...... 3 7

Figure 8: Boxplot of scaled variables across all call types (blow, bark, roar, rumble, squeak, squeal, trumpet) ...... 43

Figure 9: Scaled variables for Roar showing outliers and gradients ...... 48

Figure 10: Example histograms and boxplots of acoustic parameters for each call type show distribution and variability ...... 55

Figure 11: Final classification tree model ...... 58

Figure 12: Principal Component Analysis plot for basic call types ...... 61

Figure 13: Global ANOSIM test for difference in parameters among the call types ... 65

Figure 14: Power spectra of Rumble and background noise at the Oregon Zoo ...... 75

Figure 15: Power spectra of Rumble and background noise at Elephant Nature Park ...... 76

Figure 16: Power spectra of rumbles and background noise at the Royal Elephant Kraal ...... 77

VI CHAPTER I: INTRODUCTION Relatively little is known about the vocal repertoire of Asian elephants

(Elephas maximus). Our knowledge of acoustic communication in elephants is based primarily on research on three wild African elephant (Loxodonta spp.) populations

(Amboseli National Park in Kenya, Etosha National Park in Namibia, central African forests), three captive African elephant herds (Disney's Animal Kingdom in Lake

Buena Vista, Florida, U.S.A., Vienna Zoo in Austria, and Daphne Sheldrick's orphanage in Nairobi National Park in Kenya), one wild Asian elephant population

(Gal Oya National Park in Sri Lanka), and one captive Asian elephant herd (Oregon

Zoo in Portland, OR, USA, formerly Washington Park Zoo). Additional elephant populations are currently under study, but publications are limited.

Over 30 calls have been published for African elephants (Loxodonta spp.)

(Olson, 2004; Stoeger-Horwath et al., 2007), and nine have been published for Asian elephants (Elephas maximus) (Olson, 2004; McKay, 1973); but most are described only by context or function and not by acoustic features. Among the published African elephant calls, six infant calls have been described by both spectral and temporal features in addition to context and function, so there is advancement in our understanding of vocal ontogeny (Stoeger-Horwath et al., 2007). However, less than

16 calls produced by sub-adults and adults have been described by temporal or spectral features, and most of these appear to be variations of the low frequency rumble that differ in social context and function (Olson, 2004; Langbauer, 2000).

Among the published Asian elephant calls, only the low frequency rumble has been

1 described by spectral and temporal features. African elephants are capable of producing calls ranging from 5 Hz to 9 kHz (Poole, in prep). Infrasonic calls (below

20 Hz) have been recorded from captive Asian elephants (Payne et al., 1986), but the frequency range of Asian elephant vocalizations is unknown.

TAXONOMY AND CONSERVATION STATUS

Elephants belong to the order Proboscidea and family Elephantidae. The two extant genera of the family Elephantidae include the African elephants (Loxodonta africana, Blumenbach 1769) and the Asian elephants (Elephas maximus, Linnaeus

1758).

Subspecies taxonomy of Elephas maximus varies among authors, but the

International Union for Conservation of Nature and Natural Resources (IUCN) follows Shoshani and Eiesenberg ( 1982) in recognizing three subspecies: E. M. indicus on the Asian mainland (India and Indochina), E. m. maximus on Sri Lanka, and E. m. sumatranus on the Indonesian island of Sumatra (IUCN, 2008). These subspecies designations were based primarily on morphological differences, but mtDNA variation suggests that E. m. sumatranus is monophyletic (Fleischer et al.,

2001). Borneo's elephants are not currently listed as a subspecies and their origin is controversial; however, comparison of mtDNA of Borneo elephants to that of Asian elephants across their range suggest that Borneo's elephants are genetically distinct, with divergence indicative of a Pleistocene colonization of Borneo (Fernando et al.,

2003). Given genetic distinction, the IUCN could define these taxons as evolutionarily significant units (ESU) (IUCN, 2008).

2 The IUCN uses Sukumar's (2003) estimate of 41,410-52,345 wild elephants for the global population, but acknowledges the argument by Blake & Hedges (2004) that these estimates may be inaccurate due to the challenges of surveying forest­ dwelling populations. A more recent estimate lists a wild population of 38,534-52,566 animals (Sukumar, 2006) and a global captive population of approximately 16,000

(Sukumar, 2006; Fischer, 2004). The Asian elephant is listed as Endangered by the

IUCN (Choudhury et al., 2008. Elephas maximus. In: IUCN 2008) and is listed in

Appendix I of the Convention on International Trade in Endangered Species of Wild

Fauna and Flora (CITES) (CITES, 2008). The greatest threat to the Asian elephant is habitat loss, degradation, and fragmentation driven by an expanding human population, which leads to increasing conflicts between human and elephant. Poaching for ivory and other body parts also pose a major threat to the long-term survival of some populations. Large-scale hunting has reduced populations significantly over wide areas; but even selective removal of tusked males can result in skewed adult sex ratios and reduced genetic variation, especially in areas where populations are isolated

(Choudhury et al., 2008. Elephas maximus. In: IUCN 2008).

Although there is genetic evidence that suggests there may be at least two species of African elephants, namely the African savanna elephant (loxodonta africana) and the African forest elephant (Loxodonta cyclotis) (Roca et al., 2001), both the IUCN and CITES classify all African elephants as a single species encompassing both savanna and forest populations (IUCN, 2008; CITES, 2008). Many members of the international.elephant management and medical community currently recognize

3 the reclassification of Loxodonta africana into two separate species, Loxodonta africana and Loxodonta cyclotis, with the following subspecies of Loxodonta africana, L.a. africana (South African bush elephant), L.a. Knochenhaueri (East

African bush elephant), and L.a. oxyotis (West African bush elephant).

The African elephant is listed as Near Threatened by the IUCN (Blanc, 2008.

Loxodonta ajricana. In: IUCN 2008) and is in Appendix I of CITES, with four populations down-listed to Appendix II (CITES, 2008). Poaching for ivory and meat remains a significant threat, but the most important perceived threat is habitat loss and fragmentation caused by an expanding human population and land conversion, which in tum leads to an increase in human-elephant conflict (Blanc, 2008. Loxodonta africana. In: IUCN 2008).

ECOLOGY AND COMMUNICATION

Elephants are intelligent, long-lived animals that live in a complex and fluid society in which several modes of communication play a role in maintaining group cohesion and social order, and in locating and assessing reproductive state of potential mates (Langbauer, 2000; Eisenberg et al., 1971). Asian and African elephants have a similar family structure. Females live in matriarchal herds consisting of other female relatives and their young. Elephant society is multi-tiered, with family units joining to form bond groups or larger clans that share a home range and coordinate movements

(Douglas-Hamilton, 1972; Charif et al., 2005). Males are excluded from familial units at a young age and live as solitary elephants or in temporary association with other males in "bachelor herds" at the fringes of the female herds (Eisenberg et al., 1971;

4 Poole, 1987; Vidya & Sukumar, 2005). Recent studies suggest that bull society may be more structured than previously thought, and that young bulls may seek the company of mature bulls (Kate Evans, pers. comms). There is evidence that larger clans may not form in Asian elephant society, so ranging behavior may be influenced to a lesser degree by extended kin compared to African savannah elephants (Fowler &

Mikota, 2006).

Data on Asian elephant mating behavior in the wild are scarce and incomplete, but studies of African savannah elephants suggest that mate choice by females and mate guarding by males appear to play crucial roles in the reproductive behavior of at least the African savannah species (Moss, 1983; Poole, 1989).

Infrasonic vocalizations and chemical signaling are considered primary modalities for long-distance communication (Payne et al., 1986; Poole, 1999); whereas visual, tactile, auditory, and chemical signals are all important at close range

(Schulte & Rasmussen, 1999). Recent findings with seismic detection and discrimination suggest that vocalizations traveling in the seismic channel could be used for both short- and long-distance communication (0' Connell-Rodwell et al.,

2006; O'Connell-Rodwell et al., 2007). Acoustic signals are temporally short-lived and both sender and receiver must be present for the signal to be communicated, and thus these signals provide information only on an immediate situation, for example, location (Langbauer, 2000). Chemical signals are temporally long-lived and the sender and receiver need not both be present for the signal to be communicated, and thus ·

5 these signals provide information on a constant state, for example, reproductive state

(Langbauer, 2000).

VOCAL PRODUCTION AND INDIVIDUALITY IN ACOUSTIC SIGNALS

Mammals can modify certain basic sounds by changing the amplitude, temporal patterning, or stressing of overtones (Muckenhim in McKay, 1973). Vocal individuality has been found in other mammalian species, for example bottlenose dolphins (Tursiops truncatus) (Janik et al., 2006), rhesus macaques (Macaca mulatta)

(Rendall et al., 1998), fallow deer (Dama dama) bucks (Reby et al., 1998), swift fox

(Vulpes velox) (Darden et al., 2003), but it is unknown how or if individual identity is communicated via the acoustic channel in Asian elephants. Strnctural variation within a call may serve a communicative function (Green, 1975 & Bayart et al., 1990 in

Soltis et al., 2005), but call structure must first be characterized.

Characteristics of vocalizations that arise from inherent properties of vocal fold vibration in the larynx (the source) vary independently from properties that arise from vocal tract resonance (the filter), so either the source or filter may provide the auditory characteristics a receiver needs to determine identity (McComb et al., 2003 ). Air driven from the lungs sets the vocal folds in motion. Vibration of the vocal folds in the larynx has afundamentalfrequency and harmonics. In terrestrial mammals, the fundamental frequency is determined by how many times the vocal folds vibrate in one second, measured in cycles per second or Hertz; and the fundamental frequency is negatively correlated to vocal cord mass (Fitch, 1997). The fundamental frequency determines the pitch that we hear. Harmonics of sound waves are integer multiples of

6 the fundamental frequency; that is, if the fundamental frequency is 25 Hz, then the

harmonics have frequencies of 25 Hz, 50 Hz, 75 Hz, etc. The human auditory system

does not perceive separate harmonics in sound, but harmonics provide the tone quality

or richness of the sound.

The source sound from the larynx passes through the vocal tract, which selectively amplifies certain frequencies, and thus filters the spectral envelope to produce peaks calledformants (McComb et al., 2002). As vocal tract changes shape, its resonance frequencies change, and different formants are produced. Articulators, t.he tongue and lips, produce more complex configurations of the vocal tract and hence more complicated formant patterns (Baken, 1996). The average spacing between consecutive formants, or formant dispersion, can provide information on the length of the vocal tract in mammals (McComb et al., 2003), with an increase in vocal tract length resulting in a decrease in formant spacing in many mammalian species

(Sanvito et al., 2007; McComb et al., 2003; Fitch & Hauser, 2002).

The sound energy received is a factor of sound production intensity and acoustic propagation loss in the sound channel, which is influenced by topography, boundary conditions (surface and vegetation), and atmospheric structure (temperature, humidity, pressure) (Garstang, 2004).

The vocal tract of an elephant is the nasal passages of the skull, the trunk, and the pharyngeal pouch (Garstang, 2004). The pharyngeal pouch is at the posterior one­ third of the nasopharynx (Fowler & Mikota, 2006). The elephant tongue is unable to protrude from the mouth due to the attachment from the tip of the tongue to the floor

7 ·--1

of the mouth (Fowler & Mikota, 2006), so the tongue is limited in the articulation of

sounds.

Adult Asian elephants have a size range of 2,000-5,500 kg (4,410-12,125 lbs)

and 2-3.5 meters (6'7"-11'6") at the shoulder (Shoshani, 2000), and there is a visible

difference between individuals in the size of the head and trunk, or the vocal tract.

Center frequencies of formants depend on the size and posture of the caller, and

different individual African savannah elephants have been found to overlap in these

frequencies (McComb et al., 2002). However, there has been no comparative work

done on the size of the vocal folds of elephants.

Soltis et al. 2005 (2005) found that structural variation in the rumble from adult female African elephants housed at Disney Animal Kingdom (Lake Buena Vista,

Florida, U.S.A.) reflected individual identity and negative emotional arousal of caller.

In the presence of dominant females, subordinate females produced rumbles with low tonality and unstable pitch compared to rumbles produced outside the presence of dominant females. Results of acoustic playback studies with African savannah elephants on female recognition of the contact call (McComb et al., 2003) and on signal assessment of the musth rumble (Poole, 1999) suggest individual- and size­ related differences in acoustic production, which elicit questions regarding individualism, social recognition, and size assessment. Adult females were able to discriminate familiar and unfamiliar contact calls, and it appeared that the fundamental frequency contour extracted from the harmonics was the key characteristic in the signal that could be used for distinguishing individual calls at a distance (McComb et

8 al., 2000; McComb et al., 2003). Adult males responded to playback of musth rumbles in a way that suggests they are able to assess characteristics of the caller in this acoustic signal, but it is unclear whether they assess size in the signal or whether they recognize the caller. The musth rumble is a low-frequency vocalization produced only by males in musth, and is shown to advertise the musth state (Poole, 1987; Poole,

1989). The musth rumble has been studied extensively in the African elephant but not in the Asian elephant. In response to the musth rumble of high-ranking males, other musth males approached aggressively, while non-musth males walked away (Poole,

1999), which was consistent with prior studies of this population on the state of musth, dominance, and a willingness to contest access to females (Poole, 1989). Elephants are long-lived, intelligent animals, and they meet and interact over a period of decades, so it is reasonable to hypothesize that males also recognize individuals by their calls (Poole, 1999), and that they assess size based on prior interactions rather than assessing the size of an individual by their call. To further investigate the potential for size assessment in acoustic signals, the acoustic features that have the potential to code for size (fundamental frequency and formant dispersion) need to be measured for calls used over distances or barriers where visual access is limited.

LoW-FREQUENCYSOUNDS

lnfrasound is an anthropocentric term for describing frequencies below the human hearing range, which is nominally 20 Hz to 20 kHz. The long wavelengths of low-frequency sound are more likely to hit only large objects and be reflected, and are thus resilient to atmospheric attenuation with distance (Langbauer et al., 1991; Larom

9 et al., 1997). Low-frequency sound below 100 Hz shows little attenuation in forest environments (Marter et al., 1977; von Muggenthaler et al., 2001). fufrasonic vocalizations of elephants have fundamental frequencies typically in the range

14 Hz to 35 Hz, with harmonics that extend into the audible range (Poole et al., 1988;

Payne et al., 1986; Langbauer et al., 1989; Langbauer et al., 1991). Elephants can vocalize in the infrasonic range up to levels of 103 +/- 3 dB at 5m from the source

(Garstang et al., 1995). The role and production mechanism of infrasound varies among species. Other species known to produce infrasonic sounds include the okapi,

Bengal and Siberian tigers, giraffe, and Sumatran rhino (von Muggenthaler, 1992; von

Muggenthaler, 2000; von Muggenthaler et al., 200 I; von Muggenthaler et al., 2003 ).

The hyoid apparatus of elephants differs from the basic mammalian scheme and allows a greater flexibility of the larynx, which may aid in the production of infrasound (Langbauer, 2000). This difference in the hyoid apparatus presents a potential proximate cause for production of infrasound, and the need for long-distance communication presents an ultimate cause. McComb (2002) suggested that elephants produce fundamental frequencies in the infrasonic range simply because of their large size rather than an evolved mechanism for long-distance communication·. McComb

(2002) analyzed the source- and filter-related acoustic features of the African elephant female contact call in the Amboseli National Park population, measured degradation of the spectral structure of this call with distance, and performed playback to measure signal perception and assessment. Harmonics peaks around 115 kHz were most prominent and had the highest persistence with distances of 0.5 km to 2.5 km. Female

10 groups showed signs of detecting calls from distances of 2 km to 2.5 km, but did not respond as though they were familiar or unfamiliar until the distance narrowed to 1.0 or 1.5 km. These results suggest that although lower frequencies have the potential to travel further, the most important frequency components for long distance communication of social identity in African elephants may be in the harmonics around

115 Hz rather than in the infrasonic range of the fundamental frequency of the contact call (mean 16.8 Hz) (McComb et al., 2002).

HEARING

There is some evidence that elephants may be better adapted for perceiving acoustic signals in the 100 Hz to 4 kHz range than in the range of the lowest frequencies produced. Heffner & Heffner (1982) measured the hearing sensitivity of a captive Asian elephant, and found the elephant to have an audibility curve similar to that of other mammals, but with a greater sensitivity to low frequencies and lower sensitivity to high frequencies than any other mammal tested prior to this study. The elephant's absolute threshold was 16 Hz (at 65 dB) to 12 kHz (at 72 dB), with a 17

Hz to 10.5 kHz hearing range by the 60 dB criterion. Although the elephant had a low frequency threshold of 16 Hz, it was considerably less sensitive to frequencies below

100 Hz than to those between 100 Hz and 4 kHz, and the maximum sensitivity was

1 kHz (at 8 dB). From 31.5 Hz to 2 kHz the thresholds were close to background noise level, so it is possible that the animal's actual sensitivity in this region was masked by background noise. Frequency discrimination tests indicated the elephant's frequency discrimination was best below 1 kHz.

11 -----,

Localization of sound relies on the time of arrival to the ear and the intensity of

sound reaching the ear. Because the interaural distance in the elephant is large, sound

reaches one ear long before the other, and the intensity of the sound on the ear closest

to the sou_nd is greater than on the opposite ear due to the buffering effects of the head.

The larger the interaural distance, the greater the ability to localize low frequency

sounds because the greater interaural distance means a larger phase and amplitude and

difference of the incoming sound wave between the two ears. Heffner & Heffner

(1982) conducted localization experiments for a frequency range of 125 Hz to 8 kHz,

source angles of 0° to 60°, and various stimuli. The elephant performed best at

localizing sound below 300 Hz and was virtually unable to localize 4 kHz to 8 kHz.

Pinnae extension appeared to play a role in localizing sound in this study

(Heffner et al., 1982).

PURPOSE AND SIGNIFICANCE OF THIS THESIS RESEARCH

The purpose of this research is to contribute to the basic science of elephant

communication and to conservation efforts. The specific goals are to (1) define an

acoustic repertoire of Asian elephants based on acoustic parameters, (2) investigate

how the repertoire is used by groups and individuals, (3) compare manual and

automated classification to validate structural distinction among call types, (4) provide

a basis for comparing acoustic communication among elephant species and populations, and (5) explore the potential for using call parameters to develop an automatic detector of Asian elephant calls for acoustic monitoring applications.

12 Communication is the intended transfer of information (a signal) from a sender to either an intended or unintended receiver. Communication is intended to confer some advantage to the sender, receiver or both via natural or sexual selection. This study analyzes sounds produced by elephants, but there was no examination of an intended transfer of information by measuring the response of conspecifics, so it is not implied that these sounds are signals with communicative value. Data were collected for future investigation of the communicative function of these sounds, but this topic was not included in this analysis.

This study provides a basis for future research. A categorization of basic call types and modifications of these call types by quantitative acoustic parameters is needed to examine acoustic variability within and among call types, to examine individuality, to determine communicative function of calls via playback, to compare species and populations, and to develop rigorous call recognition algorithms for monitoring wild and managed populations.

The task of vocalization classification and speaker identification is common in bioacoustic analysis. Automated methods are particularly useful with large datasets or when the repertoire is being compared between individuals or social groups (Deecke

& Janik, 2006). Detection algorithms are used extensively in the passive acoustic monitoring of marine mammals, and there are many published classification methods and tools that could be used with other vocalization datasets. Examples of methods referenced for this current study include classification of African elephant and marine mammals vocalizations. Campbell et al. (2002) used artificial neural network to

13 identify individual female Steller sea lions (Eumetopias jubatus). Clemins et al.

(2005) developed a Hidden Markov Model based on human speech recognition techniques to automatically classify African elephant vocalizations for call type and individual.

Classification of sounds by acoustic parameters is the first step in developing call recognition algorithms for call detection and acoustic monitoring or census. Payne et al. (2003) provides evidence that elephant calling patterns can be reliable indicators of group size and composition, both of which are important for acoustic monitoring.

Calls were divided into three structures, single-caller low-frequency, single-caller high frequency, and multiple-callers low-frequency. The rate of calling increased with increasing numbers of elephants, and the distribution of these call categories changed with group composition. Vocalizations may provide another tool for mitigating human-elephant conflict, which often results in injury or death to both humans and elephants (Kemf & Santiapillai, 2000). In Sri Lanka, researchers are investigating the use of speech recognition techniques as a means of remotely identifying individual elephants as they approach crops (Doluweera et al., 2003).

A library of vocalizations from known individuals in known reproductive states will facilitate a more rigorous categorization of sounds by acoustic parameters, and will allow an examination of sources of variability. An understanding of this variability is needed to reliably determine the meaning of various calls via playback

(Langbauer, 2000). A database of acoustic communication of African savannah elephants (Loxodonta africana) is currently being developed by the Savannah

14 Elephant Vocalization Project (SEVP) in Amboseli National Park (SEVP, 2008) and by Disney's Animal Kingdom (John Lehnhardt, Joseph Soltis, pers comms). SEVP has collected more than 70 different elephant call types of African savannah elephants and linked them to observations of elephant behavior (SEVP, 2008), but the descriptions of only a subset of these calls have been published. There is no similar database known to be developed for wild or captive Asian elephants (Elephas maximus), so one goal of this current study is to contribute to an animal communications database for future research.

This study aims to provide a basis for comparisons of acoustic communication in wild and captive Asian elephants, and could potentially serve as a basis for comparisons between Asian and African elephants. These comparisons may provide insights into the ecological role such communication plays and the requirements of counterpart populations. As elephants lose habitat and find themselves under varying degrees of management, the need to understand their requirements in captivity will become even more important to the survival of the species (Riddle et al., 2003).

15 CHAPTER II: PUTATIVE CALL TYPES AND MANUAL CALL CLASSIFICATION

METHODS

Study site and study subjects

Data were collected at the Oregon Zoo in Portland, Oregon, U.S.A. and in

Thailand. Data collected at the Oregon Zoo were from a captive community of three male and four female Asian elephants. During the course of this study, the females were managed as a single herd with only temporary separations. The males were housed alone with no visual or physical access to other males; however, they did have acoustic and olfactory access given their close proximity and the movement of elephants between enclosures. Males and females were housed together during social introductions and breeding. The Oregon Zoo elephant exhibit is designed with two

2 outside yards totaling 3 I 40 m , and seven inside rooms. The Oregon Zoo is in an urban setting located near a major interstate. Sources of anthropogenic noise include highway traffic, air traffic, hydraulics, water, electric fences, zoo construction, a zoo train, and visitors.

Data collected in Thailand were from two herds of domesticated elephants, one herd of approximately 80 elephants at the Royal Elephant Kraal in urban Ayutthaya, and one herd of approximately 30 elephants at the Elephant Nature Park in the rural

Mae Taeng Valley north of Chiang Mai. The Royal Elephant Kraal is a working elephant village with adult and semi-adult bulls and cows, geriatric cows, and a nursery of 10-12 calves. The Elephant Nature Park manages injured and abused

16 animals in a semi-captive setting where elephants are allowed to roam during the day

within the grounds and chained at night in social groups. At the time of recording, there were three adult bulls, two sub-adult males, two calves, and females of varying ages. Sources of anthropogenic noise in both Thailand sites were sporadic road noise, machinery, water, and faint talking of visitors.

Acoustic data collection · Data were collected at the Oregon Zoo from February 2005 to March 2009, and consisted of over 80 hours of observations and recordings, with 56 hours of usable data. Data were collected continuously for one hour (occasionally 30 min to 2 hours) during social introductions, breeding events, temporary separations, arrival of a bull, the death of a matriarch, novel events, and routine husbandry. TQe following data were collected during each session: social context. behavior, vocalizations, and visual signs of musth. Reproductive state measured by hormone levels was provided by the Oregon

Zoo. The caller, if identified during observation by sound localization or visual cues, was dictated into the recorder or noted on a checksheet. One challenge was detecting and localizing vocalizations at low frequency; however the harmonics often extended into the audible range, making it possible to detect and localize at close proximity to the caller.

Data were collected in Thailand from November to December 2009, and consisted of approximately 6 hours of data. Data were collected continuously for one hour sessions (or opportunistically for short sessions) during the morning release to pasture, morning routines, night feedings, greetings between human and elephant, and

17 elephant painting. The following data were collected during each session: social

context, vocalizations, and focal animal if there was a focal.

Acoustic data were collected with four systems during the course of the study.

The frequency responses given for these systems are those provided by the

manufacturer, unless otherwise specified. The systems used were: (1) M-Audio

MicroTrack II (20 Hz -20 kHz,+/- 0.5 dB, recorder tested down to 10 Hz+/- 3 dB),

(2) Bruel & Kjaer 4145 condenser microphone (2.6 Hz -18 kHz+/- 2 dB), an ACO

type 012 preamp (flat to 0.5Hz), an ACO PS2000 power supply, and a Racal V-Store

24 Instrumentation Recorder (analog, DC-45.5 kHz), (3) Edirol R-09 mp3 recorder

(20 Hz - 40 kHz+/- 2 dB), (4) acoustic extraction from digital video (Panasonic PV­

DV700, 20 Hz - 20 kHz) that is being used as part of an ongoing behavioral study.

Only the M-Audio and digital video were used to collect data in Thailand. Preliminary data were collected with an loTech Wavebook 512 12-bit IMHz Data Acquisition

System, USBGear USB Sound card modified to record to DC, and Vetter 820 analog recorder.

During recording sessions, the microphone was fixed in position with a tripod to minimize extraneous noise from holding it directly, as per McComb (1996).The distance to source was variable as subjects moved throughout the recording session, so distance to source was only approximated for each recording session and there was no attempt to measure absolute intensity of calls.

A Cambridge Electronic Design (CED) Micro1401 rnkII data acquisition unit running Spike2 software (v5.12) was used to upload analog acoustic data. Adobe

18 Audition (vl.5) was used to convert file formats and split recording channels. Files

were first down-sampled to 25kHz to reduce the file size. The sampling rate was selected based on the published frequency range of African elephant calls, 5 Hz to

9 kHz (Poole, in prep) which gives a minimum sampling rate at the Nyquist frequency of 18 kHz. However, energy was found to extend above 12 kHz, so the original sampling rate of either 32 kHz or 44 kHz was used for final measurements of acoustic parameters.

Scoring calls and classifying calls into putative call types

The ethogram used for preliminary data collection included vocalizations described by McKay (1973) and those in the Elephant Husbandry Resource Guide

(Olson, 2004). The ethogram of elephant behaviors in the Elephant Husbandry

Resource Guide is a compilation of ethograms from approximately 30 publications and manuscripts, so it is a relatively comprehensive source. Ad libitum sampling was employed to incorporate vocalizations not yet described. Six basic call types were defined (trumpets, squeaks, squeals, roars, rumbles, and barks), and these call types and modifications of these call types were fit into published nomenclature where possible. Combination calls were also defined (trumpet-rumble, trumpet-roar, squeak­ rumble, squeak-squeal, roar-rumble). Call descriptions were compared to those recorded in an ongoing study in Uda Walawe National Park in southern Sri Lanka, and a joint ethogram was developed for calls that were common to the captive herd at the

Oregon Zoo and the free-ranging elephants in Sri Lanka (Shermin de Silva, pers comms.)

19 Exemplar calls were presented to elephant handlers and university students to gather descriptions of how the sounds were perceived compared to familiar non­ elephant sounds, with the aim of providing a description that would allow elephant handlers and researchers to discuss elephant calls in terms of aural cues; for example, a Squeal sounds like rubber shoes on a wood floor.

A signal or sound can be viewed as a distribution of energy in time and frequency. These distributions can be represented several ways. A spectrogram displays a three-dimensional plot that shows how the sound varies over time, with frequency on the y-axis, time on the x-axis, and the relative power (or the logarithm of relative power) at a given point in frequency and time represented as a darkness value.

A power spectrum displays average energy of the signal over a period of time, with relative power on the y-axis and frequency on the x-axis (frequency domain). A waveform displays amplitude on y-axis and time on x-axis (time domain).

Praat software ( v4.5.16, Institute of Phonetic Sciences of the University of

Amsterdam, The Netherlands) was used to annotate (score) sounds by adding boundary markers and typing call notations (call type, caller ID, quality) in an annotation field within the boundary. Calls were categorized into the putative call types using perceptual aural cues and visual inspection of spectrograms for differentiation of the following acoustic parameters: fundamental frequency contour

(start and end frequency, maximum and minimum frequency, inflection), tonality, and signal duration. Categorizing by aural cues and visual inspection is considered a manual method of classification. With the exception of blow sounds, every detectable

20 occurrence of every call type was scored. Environmental sounds were also scored for future analyses with call detection. Half of the recordings were first scored by two trained but inexperienced reviewers, then verified and corrected. Of 1243 sounds scored by inexperienced reviewers using aural cues and spectrogram inspection, 90 required correction, for an interobserver reliability of 93%. The caller (if known) was annotated for each call. A quality score of 1 to 4 was assigned to each call using methods by Cambell et al. (2002), which were based on presence/absence of sporadic sound overlap (elephant, human, environmental) and a subjective measurement of degradation by ambient noise (e.g., highway, wind, water, electricity, faint sounds of visitors, insects).

RESULTS

The ethogram of putative call types with descriptions and exemplar spectrograms is provided in Table 1. The aural cue and spectrogram descriptions were used for categorizing sounds into call types. The visual cue was used to help identify the caller. The published descriptions include only those for Asian elephants.

21 Table I: Acoustic ethogram of putative call types showing single calls, trains of calls, and call combinations. Call type and spectrogram Definition

Basic Sounds

Aurul cuf:' : Sounds like a diesel engine. motorboat. distant helicopter: PuJsatmg with a t1a1 pitch. Perceived RUMBLE as low frequency. Spnrrogram inspt'Crion: Flar fundamental frequency with narrow bandwidth. Pulsating in spectrogram associated w/ amplitude modulation in waveform. Harmonics present. Visuul me ufcaller: Mouth visibly open. forehead flutter. Puhlished Jescriprion: A resonant growl (Olson. 2004). Growl modified by resonance in trunk. used in contex1 of mild arousal (McKay. l 97J). Call with fundamental frequency in range l 4 Hz to 30 Hz. lasting 10-1:5 sec. Aural cuf:': Sounds like a lion roar with changing pitch Pitch sounds like it changes because of amplitude modulation. Perceived as low frequency. ROAR "' Spt'crrogram inspecrion: Flat or slightly modulated fundamental frequency. Amplitude modulated. with greatest amplitude somewhere during the call rather '""' than at beginning or end. Tonal. noisy. or mixed.

~ ~ 3000 Sometimes ends in rumble. Broader bandwidth than ~ rumbles. but inconsistent. Harmonics present. Jt 2000 Visuul cue of caller: Mouth wide open or widens

1000 during call. lower abdomen contracts. Puhlished description: Growl modified by amplitude

00 05 10 1 5 20 increase. used in context of long-distance contact. Used lime ~ primarily by juveniles who have been separated from their group (McKay. 1973 ). Pulsating sound. Loud growl (Olson. 2004 ). BARK "' Aural cuf:': Sounds like a short loud grunt. clearing throat. cape buffalo. chuff of cow. cheetah. Ends ahruptly with no roll off. Noisy compared to rumble. dOOO Spt!crrogram inspn-rion: Flat hmdamentaJ frequency.

6000 Short duration. Ca!J is usuaJ!y noisy. Harmonics not always visible (or difficult to see with noise). """' Visual cue of caller: Mouth open.

2000 - . ... Puhlished description: None for A~ian elephants. but may be similar to the bark African elephant calves c.~- ·?l- - "-=- (Stoeger-Horwath et al .. 2007). ' 0 1 O? 0.3 1} 4 05 06 time. s

22 Call type and spectrogram Definition

SQUEAK Aural cue: Sounds like a wet finger across tight rubber surface. audible pedestrian crossing signal. baby alligator. gibbon caJI. High-pitched sound that falls in pitch, powerful initial sound. aooo Spectrogram inspection: Descending fundamental

6000 frequency that is either straight or curved in shape. Small shon blip may precede pitch fall. Harmonics 4000 present. Modified by repetition. Visual cue of caller: Cheeks depressed. mouth open 2000 slightly or closed. tail sometimes erect. Published description: Not heard in wild (McKay. 0 0 0.2 o.3 o~ l 973). Aural cue: Train of repeated Squeaks with short breaks between Squeak sounds of about l sec. Rate of SQUEAK TRAIN squeaking may change. Spectrogram inspection: Squeaks are in a train if time between edges (end of one/beginning of next) of Squeaks is less than l sec. Ot was easier to perceive a ~ ~ 8000 break of l sec than 0.5 sec. so l sec was used to define i 4000 a train versus a bout of individual calls.) Amplitude sometimes shows decrease of subsequent Squeaks """' _towards the end of the train. Harmonics present. Visual cue of caller: Cheeks depressed, mouth open slight] y or closed. tail sometimes erect. Published description: Chirp=multiple short squeaks. Possibly in context of conflict (McKay. l 973). Sounds like rubber shoes on a gym floor, release of air SQUEAL out of a baJloon, windshield wiper on a dry window. a

10000 dog whine. Short utterance. aooo Acoustic description: Flat or modulated pitch. Short duration (less 1-J .5sec from aural cues), usually flat in pitch or difficult to tell (because it looks like just a point). Harmonics present. Modified by repetition. Visual cue of caller: Cheeks depressed, mouth open

0.IJ ., 0 4 0.6 o.e 1.0 slightly or closed. ume '> Published description: none

23 Call type and spectrogram Definition

Aural cue: Train of repeated Squeal with shon breaks SQUEAL TRAIN between sounds. Sounds like squeals in staccato. intennittent release of air from a balloon. Each squeal sounds flat in pitch but they jump up and down. 0000 Specrrogram inspecrion: Squeals are in a train if time ~ between edges (end of one/beginning of next ) of r~ Squeals is less than l sec. Flat fundamental frequency 4000 of each squeal. with frequency shifts or jumps between Squeals in the train. Harmonics present. 2000 Visual cue of cu.ller: Cheeks depressed. mouth open

IO slightly or closed. Published description: none Aural cue: Sounds like continual release of air from a SQUEAL LONG balloon. fingers on a wet baJloon. sliding pitch. sounds Jong. Specrrogram inspection: Modulated fundamental aooo frequency with many inflection points. Longer than Squeal but duration variable. clearly longer than l .5- ~ "'6000 2sec. Sound is continuous with no apparent breaks and J.... pitch slides rather than jumps. Frequency values similar to Squeal Trains but with continuous energy

2000 rather than breaks. Harmonics present. Visual cue of caller: Cheeks depressed. mouth open slightly or closed. Published description: none Aural cue: Sounds like a trumpet blast. Perceived as a higher frequency call. Specrrogram inspection: Flat fundamental frequency.

6000 Narrowband. Sometimes has broadband energy at higher frequencies (top hat of noise). Harmonics ~ "'6000 ! present. """' Visual cue of caller: maybe a lifting at base of trunk. trunk sometimes extended. tail sometimes erect 2000 sometimes (any behavior indicating extreme arousal). Published description: Squeak modified by increased amplitude and duration of call. Used in context of extreme arousal. Pulsating sound (McKay. l 973).

Aural cue: Forced exhalation through end of trunk. Not a vocalization. Specrrogram inspection: Noisy broadband. Looks like a foot scrape. but lower frequency than foot scrape. Harmonics not clearly visible. Visual cue of caller: air blast from trunk

0 0 02 0 4 0.6 08 I 0 .., Published description: May be same as snort in McKay '""' ' (1973).

24 Call type and spectrogram Definition BLOW-TONAL Aural cue: Sounds like blowing into a glass bottle or 250C rubbing water on the rim of a crystal glass. Very clear.

1000 rl Specrrogram in~pection: Tonal. Flat fundamental

~ 1000 !1 frequency. Harmonics not clearly visible. ; I lw'° ii Visual cue of caller: Forehead protrudes enough to see ""' at a distance. Produced only by one animal in study (Oregon Zoo male. Tusko). o:i os 10 15 20 2.5 30 35 '""'. Published description: none

Combination calls Calls in which constituent parts are basic calls types. but with a continuous frequency contour and no apparent break in sound between the constituent parts. The criterion of continuity differentiate combination calls produced by a single individual from overlapping calls produced by multiple individuals. ROAR/BARK + RUMBLE '"" A Roar or Bark ending in a Rumble. 3500 Same acoustic structure as singles calls. but with :woo continuous sound and frequency contour. ~ ;. ~ Ii May just be a variance of Roar. Many Roars have at ~ 2000 /! 1500 least a slight fall off at the end that sounds like a ,,.. Rumble. In (Stoeger-Horwath er al.. 2007). Roars are "'° broken down into mixed Roars of various combinations.

•0000 .,,., Any combination of Squeals and Squeaks occurring in

~ a train with less than l sec between the edges of the rff.Ol'IO ~ calls. 1- Same acoustic structure as singles calls. """'

A Squeak ending with any variation of Bark. Rumble. or Roar. Usually ends with a Bark. '""" Same acoustic structure as singles calls. bur with '""' continuous sound and frequency contour. ""' ,, ·10 ., •• .. " "

25 Call type and spectrogram Definition TRUMPET+ ROAR/RUMBLE

A Trumper endmg. in a Roar or Rumble. Trumpet . ending. in Rumble' is rare . Same acoustic structure as singles call s. bul with continuous wund and frequency contour.

:;.:; >.

26 CHAPTER III: CALL DISTRIBUTION AND CALL RATE - HOW THE REPERTOIRE IS USED BY GROUPS AND INDIVIDUALS

METHODS

Comparing call distribution of groups and individuals

To compare incidence of call types among groups and individuals, the number of calls of each putative call type (Table 1) was totaled for the study sites and for individuals in the Oregon Zoo herd. For this comparison, the following calls were removed from the dataset: duplicates of calls recorded with multiple systems, Blows, and recordings of one female elephant who vocalized on command. All combination calls with basic call types as the constituent parts were combined into one group called

"combo."

The call rate was not weighted to account for differences in recording time among sessions. For comparison of the Oregon Zoo herd to domesticated elephants in

Thailand, the recording duration, sample size of calls, and number of subjects were very different, but the contexts in both sites offered situations of mild to extreme arousal. For comparing individuals within the Oregon Zoo herd, the call rate was not weighted because the total recording time for the cows was similar. Also, the caller has not yet been identified for every call, so these call compositions provide only a snapshot of how each cow uses the repertoire. Only one male, Tusko, was included in this comparison because of limited data for the other males.

27 Comparing call rate in various social contexts The rate of calling, regardless of the caller, was compared among the various contexts in which data were collected at the Oregon Zoo. For this comparison, the following calls were removed from the dataset: duplicates of calls recorded with multiple systems, Blows, recordings of one female elephant who vocalized on command, and males recorded when by themselves. The total number of calls was

2147. The number of calls of each call type per recording session were calculated, then adjusted to a 60 minute recording duration. The social contexts in which recordings were made are shown in Table 2. Most of the recordings were scheduled during social introductions, with some opportunistic recordings.

Table 2: Social contexts for call comparison

Context Description #of sessions

Routine No specific event, normal training, husbandry, shifting 6 Enrichment Novel enrichment in yard (not just browse, maybe new tires or 3 something rare) Introduction 1 male in with female group (breeding, intro), includes howdy 39 Reunion Females join after separation of group for an Introduction event 1 after introduction Separation Focal animal is separated and by themselves (not with male as in intro) 2 Reunion Females join after a Separation event 1 after separation Transfer Arrival of bu11 Tusko 1 Death Euthanasia of matriarch Pet, including time during euthanasia, visitation 1 of herd members to the body, other herd members after leaving the body

28 RESULTS

Call distribution between sites With the exception of the Squeak call. the domesticated elephants recorded in

Thailand produced the same repertoire as the Oregon Zoo elephants. but with a different distribution of calls, as shown in Figure 1. The Squeak call was made by

Thai elephants. but only when asked to speak or when engaging tourists. Some factors that potentially account for the difference in call distribution are a difference in social context. difference in number of animals in the herd. environmental noise, and individuality.

A. B. Combo Rurrble 3% Rumble Corrbo 6%

37%

Roar 12%

ark

25%

Squeal 23%

Figure J : Call distribution of the Oregon Zoo herd versus domesticated elephants in Thailand Panel A shows the call distribution of the Oregon Zoo herd (N=2066 calls). Panel B shows the call distribution of domesticated elephants in Thailand (N=279 calls).

Call distribution among individuals

Individuals within the Oregon Zoo herd use the repertoire quite differently. as shown in Figure 2 and Figure 3. For most of the recording sessions, the cows were together as a group, so the difference in calls produced could be a function of 29 individuality. differences in roles within the herd. or differences in level of arousal within the same social context. Investigating these differences is planned with future analyses. Only one of the three bulls was recorded consistently during the course of this study, and this bull produced primarily one sound. a tonal Blow.

A. B.

Bark RuntJle RuntJle Roar 0% Trurrpe\ ContJo 1% Roar Bark ContJo 1% 1% Squeak 0% 3°/o. . 0% .'.· ·. ~

Squeal 27%

Squeak 54%

c. D.

Figure 2: Call distribution of females in the Oregon Zoo herd. Panel A shows Pet. Panel B shows Sung-Surin (Per's daughter). Panel C shows Rose Tu. Panel D shows Chendra. Panel E shows Tusko (male).

30 Run"ble 0% Roar Squeak 6% 0% Squeal 0%

2% Con"bo 0%

Blow-tonal 81 %

Figure 3: Call distribution of one male (Tusko) in the Oregon Zoo herd.

Call rate in various social contexts

Figure 4 shows a comparison of call rates as a function of social context. The highest rate was in temporary separatjons and reunion after those separatjons. Not surprising is the low rate of calling during routine husbandry. What is surprising is the relatively low rate of calling during the death of the matriarch, the transfer in of a bull. and social introductions. Given that elephants use many modes of communication, it is possible that the rate of acoustic production for this herd is not a reliable indicator of arousal, or that what we perceive as a situation that warrants increased arousal does not actually do so.

31 Call Rate by Context 70

60

.. 50 :::J 0 ;i: GI 40 'tii a: ~ 30 c Ill ~ 20

10

0

-~0 ~ .,p -o~ ~<§:' ~o~ ~~ !§' ~ ~0 ~ ~<:<) ()0 «-0 ~-s J>-'>CJ 'b-{b' . ~'lf !...°"'"" ~o c-;,0~ c-;,0~ "{b' i..'O' ,~ «,.~ '~

Figure 4: Mean call rate as a function of social context

32 CHAPTER IV: ACOUSTIC PARAMETERS AND AUTOMATIC CLASSIFICATION OF CALLS

METHODS

Measuring acoustic parameters

Praat scripts were run to make a separate sound file for each vocalizations.

Only sounds of basic call types (Bark, Roar, Rumble, Squeak, Squeals, Trumpet) and

Blows with no overlapping sporadic sounds were selected for measurement. All calls

were measured at the original sample rate of either 32 kHz or 44 kHz.

Osprey (vl.7) on MATLAB (v7.7, Mathworks, Inc.) was used to measure the

acoustic parameters in order to characterize the signal structure of each call type.

Osprey's measurement system was developed for characterizing the marine sounds in

the Macaulay Library of Natural Sounds. Osprey measurements are based primarily on

Fristrup and Watkin's (1993) AcouStat approach, with additional measurements and

modifications that use estimators of central tendency and dispersion that are robust to

outliers, namely the median, interquartile range, and quartile skewness. Measurements

that use these estimators of central tendency have more consistent values at variable

noise levels than measurements that rely on manual selection of signal extremes

(Mellinger & Bradbury, 2007; Cortopassi, 2006). Signal extremes are sensitive to

noise and outliers. In addition, manually measuring signal extremes requires

assessment of signal onset and offset in both time and frequency, which can be

affected by display settings and can be biased by researcher expectation and experience (Cortopassi, 2006).

33 Calls of the same type and sampling rate were merged together into a sequence of calls for quicker and more consistent measuring. Spacers were added between calls in the merged file for Rumbles and Squeals to help distinguish separate calls. The spectrogram parameters in Osprey were set to show separation of harmonics and good detail in both time and frequency. and were set the same for each call type (window type=Hamming, hop size=14 , zero padding=lx. frame size (digital samples per FFf)

=512 to 2048, for a filter bandwidth of 63 to 349 Hz). An annotarion box (or selection box) containing the entire sound in a region of time and frequency was drawn liberally to contain the entire sound plus some background noise, as shown in the two example calls in Figure 5 and Figure 6. Because only calls with no overlap of other sporadic sounds were used for measuring acoustic parameters. the annotation box included only the ambient noise and the focal sound.

8000 .. 6000 ::i:.>. 0 ..c :::J ..er .:: 4000

2000

0 000 0.05 0 10 015 020 025 030 0 35 0 40 0 45 050 ti me . s Figure 5: Osprey window of Squeak with annotation box and feature box. X-axis is time. Y axis is frequency. Coloration is intensity. Annotation box encompasses the entire signal. Feature box encompasses 90% of the signal energy.

34 12000

10000

:r 8000 ~ c:: " ~" 6000

4000

2000

0.2 0.4 0.6 0.8 1.0 1.2 1 4 1 8 20 time. s Figure 6: Osprey window of Trumpet with annotation box and feature box. X-axis is time. Y axis is frequency. Coloration is intensity. Annotation box encompasses the entire signal. Feature box encompasses 90% of the signal energy.

Osprey creates a spectrogram of the annotation box and applies a de-noising algorithm to the entire spectrogram to remove general ambient noise (Mellinger &

Bradbury. 2007). It then creates a feature box that encompasses the inner 90% of the signal energy in time and frequency, with the strongest 90% of the signal represented and the weakest 10% excluded. In the time domain (horizontal axis in the spectrogram), energy in each column is summed, with the sequence of sums forming a row vector known as the time envelope. The row vector is then ranked (sorted) highest to Lowest in energy. and the cumulative sum beginning at the high end is computed until 90% of the total energy is reached. The earliest and latest time indices included in this 90% cumulative sum define the time bounds of the feature box. The inner 90% is the upper 90% after the values in the time envelope are sorted into ascending order.

An analogous process happens in the frequency domain (vertical axis in the

35 spectrogram), with the sequence of row sums forming the frequency envelope (power spectrum), and the 90th percentile of the sorted energies being used to define the frequency bounds of the feature box.

After the feature box is established, Osprey extracts 28 acoustic features within this feature box and calculates the signal to noise ratio within the annotation box.

Osprey weights measurements at each instant by the normalized intensity of the signal at that instant. This weighting means that louder parts, which are least affected by background noise, have the strongest influence on the measurement value (Mellinger

& Bradbury, 2007).

For calls that have low frequency energy around 20 Hz (Rumbles, Barks, and

Roars), the frequency of the fundamental was verified by measuring the harmonic spacing of the call in Praat. The annotation box was then drawn so the boundary matched the fundamental frequency rather than drawing it more liberally to include energy below the fundamental frequency. The reason for this was twofold: (1) one goal of this study was to determine the frequency range of elephant calls, so an accurate measure of the minimum frequency for low frequency calls was needed, and

(2) drawing the annotation box below the fundamental frequency sometimes resulted in the inner 90% of the signal energy encompassing energy below the frequency that was verified as the fundamental frequency by harmonic spacing.

36 Acoustic parameter definitions

The acoustic parameters measured by Osprey are illustrated in Figure 7 and described in Table 5. A complete descrjption of the calculation of these parameters is given by Mellinger & Bradbury (2007).

A.

10000

8000

...

~:>. 6000 u i :;J e- - 4000

2000 M11JM20

0

M17 M7 time. s B.

i~)

1· 1 . M25 v-J i\ · ~ ID \ 1 V\ M27 l. I) I • ~ ~J \ _: · M6 f\ g / \ ~ _ii F'I ' \. /\-//\r_J '- ·-J, ~ J ,_,, \t'h (' u~ I I

Ml-M6: Feature box

Start Time sec Lowest time index in bounds that encompass the temporal (Ml) inner 90% of the signal strength in the time envelope.

End Time sec Highest time index in bounds that encompass the temporal (M2) inner 90% of the signal strength in the time envelope. Lower Frequency Hz Lowest frequency index in bounds that encompass frequency (M3) the inner 90% of the signal strength in the frequency envelope.

Upper Frequency Hz Highest frequency index in bounds that encompass frequency (M4) the inner 90% of the signal strength in the frequency envelope.

Duration sec Width of feature box: M2-Ml temporal (M5)

Bandwidth Hz Height of feature box: M4-M3 frequency (M6)

M7-M14: Central values and variation Uses measures that do not assume normality: median, quartile ranges, quartile skewness, concentration.

Median Time sec Time at which 50% cumulative signal energy is temporal (M7) reached. (Measured relative to start of file, so M7 was calculated as M7new=M7-Ml) Temporal Interquartile sec Concentration of a call around the median time (M7) temporal Range measured as the duration of the interquartile range of (M8) signal energy (Q3-Ql). Counts energy going forward and back from the median time (M7). Q3=median + 25% of signal energy Q l=median-25% of signal energy Temporal sec Concentration of a call measured as the time span temporal Concentration encompassing loudest 50% of time envelope values. (M9) Counts energy from the loudest parts down towards the smallest parts regardless of where the parts occur in time.

Temporal Asymmetry none Skewness of energy along time axis within temporal (MlO) interquartile range (-1.0 to 1.0)

38 Parameter Units Description Type

Median Frequency Hz Frequency at with 50% cumulative signal energy is frequency (Mll) reached. More stable than extreme values of LowerFreq and UpperFreq in varying noise conditions. Spectral Interquartile Hz Concentration of a call around the median frequency frequency Range (M 11) measured as frequency range of interquartile (M12) range of signal energy (Q3-Q 1). Counts energy going forward and back from the median frequency (Ml I). Q3=median + 25% of signal energy Ql=median-25% of signal energy

Spectral Concentration Hz Concentration of a call measured as the frequency frequency (M13) span encompassing loudest 50% of frequency envelope values. Counts energy from the loudest parts down towards the smallest parts regardless of where the parts occur in time.

Frequency Asymmetry none Skewness of energy along frequency axis within frequency (Ml4) interquartile range (-1.0 to 1.0)

M15-M20: Peak intensity

Time of Peak Cell sec Time of single loudest spectrogram cell. temporal Intensity Time of the cell containing the peak intensity. (Ml5) (Measured relative to start of file, so MIS was calculated as M15new=M15-Ml)

Relative Time of Peak % Relative time of peak intensity (M15/M5) temporal Cell Intensity (M16)

Time of Peak Overall sec Largest value in time envelope, which is the largest temporal Intensity vertical sum of the spectrogram over all frequencies. (M17) Time of the peak intensity in the trimmed time envelope. (Measured relative to start of file, so M7 was calculated as Ml 7new=Ml 7-Ml)

Relative Time of Peak % Relative time of peak intensity (M 17 /M5) temporal Overall Intensity (M18)

Frequency of Peak Hz Frequency of cell containing the peak intensity. frequency Cell Intensity (M19)

Frequency of Peak Hz Frequency of peak intensity in the trimmed frequency Overall Intensity frequency envelope. (M20)

39 Parameter Units Description Type

M21-M24: Amplitude and frequency modulation (variation of amplitude and frequency over time)

AM Rate Hz Dominant rate of amplitude modulation. amplitude (M21) Frequency of the maximum rate in the power spectrum of the trimmed time envelope.

AM Rate Variation Hz Variability of amplitude modulation measured as the amplitude (M22) width of peak at M21-6 dB. Values are discretized because at 6dB down from the peak, the widths may be a only a few bins wide so the values are integer multiples of the bin width.

FM Rate . Hz Dominant rate of frequency modulation. frequency (M23) Frequency of the maximum rate in the power spectrum of the trimmed frequency envelope.

FM Rate Variation Hz Variability of frequency modulation measured as the frequency (M24) width of peak at M23-6 dB. (How much the rate of change varies, may be related to inflections and steepness of upsweeps and downsweeps)

M25-M28: Fine features of harmonic structure, shifts in periodicity, direction of frequency change, rate of change in frequency

Cepstrum Peak Width Hz Harmonic structure structure (M25) Average width of peaks (harmonics) in power spectrum. Peak width is measured at 6 dB down from maximum value. At 6 dB down from the peak, the widths may be a only a few bins wide (like M22 and M24), but M25 is an average of integers so the values are not discretized. Narrow peaks means narrowband/tonal harmonics. Overall Entropy Hz Entropy, shifts in periodicity structure (M26) Distribution of energy across frequency blocks in a given time block. Shift from periodicity and linearity to chaos. Change in noisiness v. tonality.

Upsweep Mean Hz Direction of frequency change frequency (M27) Measures how much the frequency increases. Average change in median frequency between successive time blocks, weighted by total energy in the block. Inflection points with rising and falling frequencies throughout call result in a low M28 (closer to 0) compared to a consistent directional change. Measure is weighted to emphasize contribution oflouder signal components. M27<0 means frequency is decreasing.

40 Parameter Units Description Type

Upsweep Fraction % Rate of directional frequency change frequency (M28) Counts the number of times the frequency content increases. Fraction of time in which the median frequency in one block is greater than in the preceding block, weighted by total energy in the block. Indicates how much of the call has a directional change in the frequency. Inflection points with rising and falling frequencies throughout call result in a high M28, just as a consistent directional change. Measure is weighted to emphasize contribution of louder signal components. M28 always positive.

M29: Signal strength

Signal-to-Noise Ratio dB Signal to noise ratio within the annotation box. amplitude (M29) Ratio of the signal power (loudest cell) to the noise power (power of cell at 25th percentile). Cells are ranked low to high and the cell at the 25th percentile represents noise. (25th percentile is used because the animal call likely takes up less than 75% of the total spectrogram cells.) Measurement assumes that the within the annotation box at least 25% of the cells are without a focal signal.

Dataset for measuring acoustic parameters

The original dataset contained 2791 calls representing all call types, including combination calls. Acoustic parameters were measured for basic call types (Bark,

Roar, Rumble, Squeak, Squeals, Trumpet) and Blows having no overlapping sporadic sounds and no echo (N=lOl 1). Calls that met the following criteria were removed from this dataset for statistical analysis: duplicate calls using different recording systems (N=83) and calls with a low signal-to-noise ratio (SNR< 10 dB) (N=l). The final dataset used for statistical analysis is shown in Table 4. It is important to note that the number of samples per call type was not balanced, so the statistical power is

41 low. Also, Blow sounds were not scored for every occurrence, so the sample size was very low relative to the number of occurrence.

Table 4: Calls in dataset for statistical analysis (N=927)

Call type Oregon Zoo (N) Thailand (N) Total (N) Blow 73 5 78 Bark 55 8 63 Roar 98 10 108 Rumble 38 21 59 Squeak 193 5 198 Squeal 211 31 242 (includes Squeal, Squeal train, and Squeal long,) Trumpet 123 56 179 TOTAL 791 136 927

Process for statistical analysis for call classification

The goal of the statistical analysis was to determine if the data support automatic classification of calls that were previously categorized using perceptual aural cues and visual inspection of spectrograms, and to determine which acoustic parameters differentiate the call types. The process for the analysis was as follows:

1) screen data to investigate distributions and variability, 2) standardize data to account for different units of measure, 3) determine parameters to use for statistical analysis, 4) remove outliers, 5) determine parametric or non-parametric statistical analyses, and 6) run statistical analysis for automatic classification. All statistics were run using R (v2.8.1).

Screening data for statistical analysis

Histograms and boxplots were created to determine the distribution and variability among call types, to do a preliminary assessment of variables that

42 differentiate call types, and to learn more about how these variables reflect perceptual aural cues.

Standardizing the data for statistical analysis The acoustic parameters are measured in different units, so the data must be standardized, or scaled, for the variables to be dimensionally homogeneous. A boxplot of all variables was created for the scaled dataset, as shown in Figure 8.

0 :ro 0

0 0 15

0 0 0 0 0 0 0 0 ;o 0 0 0 0 0 0 0 8 0 0 0 0 0 ° @ 0 0 o e 0 o 0 0 8 ~ i § @ .o 0 0 0 0 "'Q) ~ > -0 Q) (;3 T <.) T' J I : ' I T I I ! • rf) I aa~$s8¢fI' II ~ e$1~~a~sss~!~1e~~$' I ~ ~ i Ii ' . . . . t i i • • i • I t

0 . 0 -6 0 0 0

J 9 I I I I I I I I I I I I I I I ... 0 ~ ., ~ N ,., ~ ::; ~ ~ ~ "' "' "' "' "' fl "' i .,. ~ ~ t: 1: i ~ ::; ::; ~ ::; i i ::; ~ ~ i ~ ~ e !! "''!! E E tr g E ~ ~ ...... ~ ,:: .. !! !! 'j ~ ~ j ti I ~ g " ~ ..0: ~ 8 ;. ~ tJ ~ 'li 'li ~ .. 0 8.,. ~ 8 i ~ ...,,_ .. ... 0 ~ a ,;; 8 i if. 'i it if. it "' ~ ~ ~ ! .s !" " .l! '6 ...."' .. .. -g I! ,,. u 0 0 ~ ~ l :!l: l ... E t t J ..."" "' ::;"' (j !"' !:."' ~ " ::; ... 0 ~ ... :::> :::> ~ " t Figure 8: Boxplot of scaled variables across all call types (blow, bark, roar, rumble, squeak, squeal, trumpet) Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis.

43 Removing parameters from statistical analysis Some parameters were removed from the statistical analysis because they were

highly correlated with other parameters for this dataset (6 parameters) or were

absolute measures that could be represented by relative measures (4 parameters).

A correlation matrix was created using the scaled dataset of the basic call types

(Bark, Roar, Rumble, Squeak, Squeal, Trumpet) and Blow to determine potential

candidates for removal from further analyses. The correlation matrix measures the

strength and direction of the linear relationship between pair wise combinations of

variables (given by the Pearson correlation coefficient). Pairs of acoustic parameters

having high correlation coefficients were considered candidates for removal. Because

a correlation measures only the strength of linear relationships, it is possible that

additional pair wise combinations had a strong relationship, but the relationship was

not linear.

Correlations were greater than 0.5 for 14 pairs of variables. Although high

correlations could indicate redundancy and therefore suggest candidates for removal

from further analyses, not all candidates were removed because some relationships

may be biologically relevant for investigating acoustic communication of this species

as compared to others. For example, Overall Entropy (M26) was highly positively

correlated with three variables related to frequency range, Upper Frequency (M4 ),

Bandwidth (M6), and Median Frequency (M 11 ), so M26 was a candidate for removal.

However, the relationship of entropy to frequency range may provide insight into sound production and perception. Frequency of Peak Overall Intensity (M20) was highly correlated with Low Frequency (M3 ), so M20 was a candidate for removal. 44 However, this relationship confirms that the peak intensity occurs at or close to the fundamental frequency of the call. Other high correlations were related to central values and variation and their relationship to frequency and duration. These correlations may indicate greater stereotypy in elephant calls compared to calls of other species, so the relationship between these variables may be of biological interest.

Further analysis may be needed to determine measurements that are most important given the study subjects and recording environment.

A total of 18 acoustic parameters were used in the statistical analyses, which are shown in Table 5 with notation for each parameter that appears in the data output.

- Parameter Notation in Output Description

Lower Frequency (Hz) LowerFreqM3 Lowest frequency index in bounds that (M3) encompass the inner 90% of the signal strength in the frequency envelope.

Upper Frequency (Hz) UpperFreqM4 Highest frequency index in bounds that (M4) encompass the inner 90% of the signal strength in the frequency envelope.

Duration (sec) DurationM5 Width of feature box: M2-Ml (MS) - Bandwidth (Hz) BandwidthM6 Height of feature box: M4-M3 (M6)

Median Time (sec) MedianTimeM7 Time at which 50% cumulative signal energy is (M7) reached. (Measured relative to start of file, so M7 was calculated as M7new=M7-Ml)

Temporal Concentration TimeConcentM9 Concentration of a call is measured as the time (sec) span encompassing loudest 50% of time (M9) envelope values. Counts energy from the loudest parts down towards the smallest parts regardless of where the parts occur in time.

Temporal Asymmetry TimeAsymmM 10 Skewness of energy along time axis within (MIO) interquartile range (-1.0 to 1.0)

45 Parameter Notation in Output Description

Median Frequency (Hz) MedianFreqMl 1 Frequency at with 50% cumulative signal (M11) energy is reached. More stable than extreme values of LowerFreq and UpperFreq in varying noise conditions.

Frequency Asymmetry FreqAsymmMl 4 Skewness of energy along frequency axis within (M14) interquartile range (-1.0 to 1.0)

Relative Time of Peak PkOverallRelTM 18 Relative time of peak intensity (Ml 7 /MS) Overall Intensity (%) (M18)

Frequency of Peak Pk0verallFM20 Frequency of peak intensity in the trimmed Overall Intensity (Hz) frequency envelope. (M20)

AM Rate (Hz) AMRateM21 Dominant rate of amplitude modulation. (M21) Frequency of the maximum rate in the power spectrum of the trimmed time envelope.

AM Rate Variation (Hz) AMRate V arM22 Variability of amplitude modulation measured (M22) as the width of peak at M21-6 dB. Values are discretized because at 6dB down from the peak, the widths may be a only a few bins wide so the values are integer multiples of the bin width.

FM Rate (Hz) FMRateM23 Dominant rate of frequency modulation. (M23) Frequency of the maximum rate in the power spectrum of the trimmed frequency envelope.

FM Rate Variation (Hz) FMRateV arM24 Variability of frequency modulation measured (M24) as the width of peak at M23-6 dB. (How much the rate of change varies, may be related to inflections and steepness of upsweeps and downsweeps)

Overall Entropy (Hz) EntropyM26 Entropy, shifts in periodicity (M26) Distribution of energy across frequency blocks in a given time block. Shift from periodicity and linearity to chaos. Change in noisiness v. tonality. Upsweep Mean (Hz) UpswpMeanM27 Direction of frequency change (M27) Measures how much the frequency increases. Average change in median frequency between successive time blocks, weighted by total energy in the block. Inflection points with rising and falling frequencies throughout call result in a low M28 (closer to 0) compared to a consistent directional change. Measure is weighted to emphasize contribution of louder signal components. M27<0 means frequency is decreasing. 46 Parameter Notation in Output Description

Upsweep Fraction (Hz) UpswpFracM28 Rate of directional frequency change (M28) Counts the number of times the frequency content increases. Fraction of time in which the median frequency in one block is greater than in the preceding block, weighted by total energy in the block. Indicates how much of the call has a directional change in the frequency. Inflection points with rising and falling frequencies throughout call result in a high M28, just as a consistent directional change. Measure is weighted to emphasize contribution of louder signal components. M28 always positive.

Removing outliers

Boxplots of all variables in the scaled dataset were created for each call type to determine potential outliers within each call type. The criteria for determining outliers was very conservative in order to preserve data that may have biological relevance.

Given that the gradients of variability may have biological relevance, data points were considered outliers only if they deviated from the overall pattern of distribution and did not appear to belong to a long tail (gradient), as shown in Figure 9. Calls were considered outliers only if they were multivaried outliers. Because variance across two variables could be a result of covariance alone, calls were considered outliers and removed from the dataset only if they were outliers across at least three variables.

Because calls recorded in Thailand could represent variability in call structure not found in the Oregon Zoo call repertoire, and the Oregon Zoo data is better represented in the dataset, outlier Thailand calls were removed only if they sounded like they did not meet the definition of the calls as established by aural cues.

47 0 8

0 outliers 6 0 0 0 not outliers

0 0 0 4 0 8 0 0 0 0 0 0 0 0 0 0

0 \:) 0 0 8 Q 0 0 Q 0 I 0 0 CJ'> ow; -, 0 ~ ~ I I 0 I 8 I -, 0 g 2 I 1-r"T" I 0 I I 0 -, I I I I I 0 ~ I -, 0 I -, I I I I I 0 I > ] c<:I u (/J ()

A~8~~~taQB~QO~cig~~_J_~-'-~ .....1....1 I I _LQ I : ~ ~ 0 : I I -2 ~ 0 ~

0 0 Q

-4 -I 8 8 I I I c ... 0 ;;;. M ;:. "'::;; ..::;; "'::!: ~ ::;:"' "' "'N ..,.. "' "'N r:r c ~ i ::!: ::;: ::;: ::;:"' ::;: ::;:"' ~ "' ::;: E"' E !! '§ ~ t: .,- ,_ U- .. ~ ~ ~ "- ;:: ..u § E ;; '!;j ~ ;IS c. .. E" ~ c: "- [ a:: 2 a:: e .. U- ~ lo "O 0 ~ c ::;: 3 c. 0 "c: u ~ ~ ::;: ~ !:I. "- 0. " ""' ~ t -e 0 if"' a:: "- ai .9 ;;; "' "i -a"'

Out of 927 calls, only 19 calls (Blow N=6, Bark N=4, Roar N=3, Rumble N=2,

Squeak N=l, Squeal N=3, Trumpet N=O) met the conservative criteria for outliers.

Table 6 shows the final dataset with outliers removed (N=908).

48 Table 6: Calls in dataset (N=908) after outliers removed

Call type Oregon Zoo (N) Thailand (N) Total (N)

Blow 68 4 72 Bark , 51 8 59 Roar 97 8 105 Rumble 36 21 57 Squeak 193 4 198 Squeal 209 30 239 (includes Squeal, Squeal train, and Squeal long,) Trumpet 123 56 179 TOTAL 777 131 908

Determining parametric or non-parametric statistical analysis

After scaling and removal of outliers, normal probability plots were created for

the scaled dataset (N=908) to determine how well the normal distribution describes the

data. The normal probability plot is a quantile-quantile (Q-Q) plot against the standard

normal distribution. The observed data is ranked and the quantile is calculated. If the

observed data set matches the theoretical distribution, the shape of the plot will be a straight line where y = x.

Based on inspection of the Q-Q plots, approximately half of the variables were normally distributed and only slightly skewed, and half were either very skewed or not normal. Therefore, this dataset does not meet the assumptions of normality, linearity, and constant variance. By not meeting these assumptions, only non-parametric statistics can be used.

49 Running statistical analyses for automatic classification A classification tree was run with the original data (non-standardized) as a

preliminary analysis for determining the most important predictors for classification

into the basic call types (Bark, Roar, Rumble, Squeak, Squeals, Trumpet), and to

explore its potential as a predictive model for classifying new calls into the pre­

defined call types.

A Principal Component Analysis (PCA) of the scaled dataset was run to

determine the parameters that explain most of the variance among call types. The main

objective of PCA is to reduce the dimensionality of the dataset. Ideally, PCA should

be multinormal and reasonably unskewed. Although the data did not meet these

assumptions, the clustering of data along principal component axes can still provide

insight into the variables that group and separate call types.

Global and pair-wise Analysis of Similarity (ANOS IM) tests of the scaled

dataset were run to determine if there were significant differences in the acoustic

parameters among the pre-defined call types and between pairs of call types. If two

call types are significantly different in their parameter values, then the dissimilarities

between the call types will be greater than those within each call type. In order to

evaluate the dissimilarity within and between call types as a measure of true distance between parameter values, the Ward's linkage method was used with the Euclidian distance. The number of iterations for assessing significance was 1000. A Bonferroni correction was calculated to reduce the family-wise Type I error from 15% to 4.8%.

Other analyses considered were a Discriminant Function Analysis (DFA) and unsupervised clustering. DFA is better than PCA for discriminating existing groups 50 and classifying into pre-defined categories; however, even the scaled data did not meet the assumption of multinormality required by DFA. A hierarchical agglomerative cluster analysis of the scaled dataset was run on the Oregon Zoo data using R.

Although there appeared to be at least a weak group structure in support of the pre­ defined call types, the output plot was illegible given the number of data points so it was not possible to determine other potential call groupings from inspection.

Unsupervised clustering should be considered for future comparison of call classification methods.

The Blow sound was included in the distributions, but not the classification tree, PCA, or ANOSIM tests. This non-vocal sound is made frequent~y by both males and females and in most situations, so characterizing this sound by its acoustic structure may be useful for acoustic monitoring. However, preliminary runs of the classification tree and PCA showed that it complicated the classification and grouping of the vocal sounds, so it was handled separately from the basic call types.

Comparing low frequency calls to background noise

Although low frequency vocalizations have the potential for communication in forested environments and over long distances, they serve a communicative function only if they can be detected by a receiver. Power spectra of low-frequency Rumble calls and background noise were compared to investigate bandwidth and intensity of

Rumbles in relation to spectral characteristics and intensity of the background noise.

Only data collected with equipment capable of recording below 20 Hz were used. Paired Rumble and noise samples were selected from the same recording to

51 ensure that the recording levels were equal. Noise segments were selected as close as

possible to the Rumble to take into account that a caller may adjust the intensity of

sound production depending on background noise. The noise segments were selected

with a duration approximately equal to the rumble. A frequency range of 14 Hz to

1.2 kHz was used to compare bandpass characteristics because rumbles have a range

of 14 Hz to 1.2 kHz for the inner 90% of the call energy. Thailand recordings were

separated by site because one site is rural (Elephant Nature Park) and the other is

urban (Royal Elephant Kraal).

Osprey uses an arbitrary reference to generate power spectra, where a sample

value of 1 is effectively 0 dB. Hence, comparisons of signal and noise can be made

only if the recording levels are equal so that the power is referenced to the same value.

With this reference to the same value, the signal to noise ratio at any frequency value

is the difference of the signal power (dB) and noise power (dB). To make these comparison, the signal and noise were not separated within each sound segment; rather, the spectrums for the paired Rumble and noise segments were compared

visually to estimate the difference between the signal power and the noise power at

various frequency values.

The sample size used in these analysis was small (Oregon Zoo N=3, Elephant

Nature Park N=4, Royal Elephant Kraal N=3), so this does not constitute a complete noise analysis, but it does offer some insight into the potential of Rumble calls to be buried in noise in various recording environments. Future analyses include a more

52 quantitative analysis of the signal to noise ratio at various frequency values for an increased number of call and noise samples.

RESULTS

Call types

Six basic vocalization types were identified as structurally different, and were labeled Bark, Roar, Rumble, Squeak, Squeal, Trumpet. An additional non-vocal

sound, Blow, was also found.

Frequency range, bandwidth, and duration of acoustic repertoire

The frequency range of the basic call types (Bark, Roar, Rumble, Squeak,

Squeal, Trumpet) and Blow encompassing all of the signal energy was 14 Hz to

18 kHz. The frequency range encompassing only 90% of the energy was 14 Hz to

9 kHz. The bandwidth encompassing 90% of the signal energy for individual calls ranged from 54 Hz to 9 kHz (median 1680 Hz). The duration encompassing 90% of the signal energy varied from 0.1 sec to 14 seconds (median 0.7 sec).

Variability of acoustic parameters among call types

Only a subset of the graphical results are included here. All plots are provided in the appendices. By comparing the histograms and boxplots, it appears that the distribution of each variable differs among call types, as shown in Figure 10. From the distributions of commonly used measurements (e.g., fundamental frequency and duration), one can see how the acoustic parameters reflect aural perception. For example, the fundamental frequency, as measured by Lower Frequency (M3), is

53 consistent or stereotypic for Bark (BRK), Roar (ROR), Rumble (RUM) and Squeak

(SQK), but varies more widely for Trumpet (TMP) and even more so for Squeal

(SQL). Blow shows the greatest variance, but this may be because it is a non-vocal sound that is mostly broadband noise. Consistent with perception by aural cues, Bark,

Roar, and Rumble are lower in frequency than Squeak, Squeal, and Trumpet. Most of the energy in Rumbles is low, whereas Roars have energy that extends higher. The

Duration (M5) is stereotypic for Trumpet, Squeak, Blow, and Bark, but his highly variable for Rumble and Squeal. Squeal include the longest calls, followed by Rumble then Roar.

54 A. B.

''°.. 3000

2500 '°1~~~1100' ~ .. i 2000 --,-- ~ ~ "1" "" 1500 .. ~ ! I : _J 1000 ~] 61.W BRK ,-~ 100 500 ci::J ~ ··r 0 ~ ~ r

BRK ROR RUM SOK SOL TMP Gal

~~,.JL1tn-o 5(10 1000 1500 2000 0 500 1000 \500 21:){10 d!a.6Slowe

TMP

80 100L[60

• I 15 1 """ '°" I ""' !{) :s ~ ~ 10 ~ 0 0 :z:; .l. c ~ ' .. :J " i: 0 5 ' ~ Q_" ' ff ' l~IL_!.BRK ROR q::i $ ~ ..L 8 ~1- BRK ROR RUM SOK SOL TMP '~1LJl_/L_1 Call 0 5 10 15 o s •o 1s dla.6SOurationM5

Figure 10: Example histograms and boxplots of acoustic parameters for each call type show distribution and variability. TMP=Trumpet, RUM=Rumble, SQK=Squeak, SQL=Squeal, BLW=Blow, BRK=Bark, ROR=Roar. Panel A and B show the Low Frequency (M3) parameter values or each call type. Panel C and D show the Duration (M5) parameter values for each call type.

55 Classification tree A classification tree was run on the non-standardized data (N=836), with

outliers removed and excluding Blows. The tree served as a preliminary analysis for

. determining the most important predictors for classification, and may have potential as

a predictive model for classifying new calls into the pre-defined call types (Bark,

Roar, Rumble, Squeak, Squeal, Trumpet). The final classification tree model is shown

in Figure 10. Tree pruning was used as cross-validation for determining the final

model. The goal of cross-validation was to evaluate alternative models that reduce the

number of branches, but when the number of branches was reduced , two call types

(Bark and Rumble) were left out and the misclassification rate doubled. The final

model explained 78.4% of the variation among call types using decision rules with

only six variables: Lower Frequency (M3), Upper Frequency (M4), Duration (M5),

Frequency of Peak Overall Intensity (M20), FM Rate Variation (M24), and Upsweep

Fraction (M28).

The overall misclassification rate was 12.2%, which means successful

classification of 87 .8% of the calls into the predicted call type. Table 7 and Table 8

show the confusion matrix and the classification rate for each call type. The Squeak

had the highest successful classification rate (95 .9% ), followed by Squeal (91.2% ),

Trumpet (91.1 % ), Rumble (78.9% ), Roar (73.3% ), then Bark (71.1 % ). Rumbles were

most often misclassified as Roars. Roars were most often misclassified as Rumbles.

Barks were most often misclassified as Trumpets. The sample sizes of Bark and

56 Rumble were low, which may impact the decision rules and thus successful

classification of these call types.

The tree was most successful in classifying the group of higher frequency calls

(Squeak, Squeal, and Trumpet), but classifying Squeaks and Squeals required multiple

decision rules. Squeals were best differentiated by having a Lower Frequency (M3)

above 675 Hz, but for those Squeal calls that did not meet this criteria, it appears that

Upper Frequency (M4), Duration (M5), and FM Rate Variation (M24) differentiated

this call type. The Squeal call type is highly variable, and it appears from the decision

rules that misclassifications are primarily with the short Squeal call, which are often

low intensity compared to Squeal trains and long Squeals. Squeaks were best

differentiated by a low value for FM Rate Variation (M24), which means the

frequency modulation rate was relatively consistent throughout the call. For those

Squeaks that did not meet this criteria, it appears that Upper Frequency (M4),

Upsweep Fraction (M28); Duration, and Lower Frequency (M3) differentiated this call

type.

The group of lower frequency calls (Bark, Roar, and Rumble) appears to be

primarily differentiated by low values of Lower Frequency (M3) and Upper Frequency

(M4), then by Duration (M5). Among these call types, the Bark is separated from the

Roar and Rumble by its short duration. The Roar and Rumble are separated by the

Frequency of Peak Overall Intensity (M20), with the Rumble being below 170 Hz.

Given that Roars and Rumbles were misclassified most often as each other, this single parameter may not be sufficient for separating these call types.

57 r ,_~.,.,

·--·------~v.i...,.. ,, I ___ l O/Oll'lfl/211119 I I I -·-----·--~~Fr~M<.:_8!>!____.______sL !IQIOll6t/610

DutlltiOOi < 0 8675

LO~fll~l<325 Pk0.11tall~>"' 170 1 t2/.ll1i!il13!16J --~-~~---., JK f~24>"0011 s& ;&. lR Rl.I 42/11/0IOf1/1 OJOl(l/191'0.'0 l)IWQ,"()/311 Gi11'11!\li21Z 411314Si()l()IO

Figure 11: Final classification tree model. (TMP=Trumpet, RUM=Rumble, SQK=Squeak, SQL=Squeal, BLW=Blow, BRK=Bark, ROR=Roar) Vt 00 Table 7: Confusion matrix for tree Diagonal cells (in grey) show correct classification into the predicted class. Off-diagonal cells show misclassifications. Columns represent the distribution of calls into each leaf (call type). Rows --r------.-- ,------. .,.. __ ,,_ Bark Roar Rumble Squeak Squeal Trumpet Bark class contains 42 barks, Bark 42 11 0 0 1 1 11 roars, 1 squeal, 1 trumpet Roar 0 77 11 0 2 2 Rumble 4 13 45 0 0 0 Squeak 1 0 0 189 5 3 Squeal 0 0 0 3 218 10 Trumpet 12 4 1 5 13 163 I•~ 59 105 57 197 239 179 # calls of each call type

59 bark calls recognized by tree as bark (42), rumble (4), squeak (1) trumpet (12)

Table 8: Classification rate ( % ) for tree Diagonal cells (in grey) show correct classification into the predicted class. Off-diagonal cells show misclassifications Bark Roar Rumble Squeak Squeal Trumpet Bark 71.1 10.5 0 0 0.4 0.6 Roar 0 73.3 19.3 0 0.8 1.2 Rumble 6.7 12.4 78.9 0 0 0 Squeak 1.7 0 0 95.9 2.1 1.7 Squeal 0 0 0 1.3 91.2 5.6 Trumpet 20.3 3.8 1.8 2.1 5.4 91.1

t Bark calls classified correctly at 71.1 % with a 28.9% misclassification rate

Parameters measured by Osprey are numerous, and one could limit the parameters used in the model to those commonly measured by spectrographic analysis tools to design a tree that could be used widely by researchers of Asian elephant calls.

Principal Component Analysis

A Principal Component Analysis (PCA) of the scaled dataset (N=836) with outliers removed and excluding Blows was run to determine the parameters that account for the variance within and among call types. In accordance with broken.stick validation, which determines whether the observed pattern is significantly different from a random pattern, only PCl and PC2 should be kept. PCl and PC2 explained

59 -----ii

only 42.8% of the variance, with PCl explaining 24.6% of the variance and PC2 explaining 18.2%.

The PCA plot in Figure 12 shows some support for the grouping of call types, but with mixing of call types. The relative contribution of each variable to the principle components, or loading of principal components, highlights a subset of the variables as being important for separating call types and grouping calls within each type. The sign of the numerical loading value indicates whether the value of that parameters increases or decreases along the principal component axes. The loading is expressed in the eigenvectors of PC 1 and PC2 below:

PCl= (-0.29)LowerFreqM3 + (-0.42)UpperFreqM4 + (-0.38)BandwidthM6 + (-0.44)MedianFreqM11 + (-0.38)Pk0veral1FM20 + (0.23)AMRateVarM22 + (0.21)FMRateVarM24 + (-0.38)EntropyM26

PC2= (-0.14)LowerFreqM3 + (-0.52)DurationM5 + (-0.51)MedianTimeM7 + (-0.51)TimeConcentM9 + (-0.10)Pk0verallFM20 + (0.24)AMRateM21 + (0.24)FMRateM23 + (-0.19)UpswpFracM28

60 ..!)

(~

~ 0 I

' f"~:r~· ~' '..(' ~ - ,,.-~· Oo 0 ~-~---r 0 c 0 c• ~ .. c 0 ( • ""c ~ · ~ . . Q_ "' ~;' <> BRK -+ . ROR 0 RUM SOK 0 _J I SOL TMP

i.J.- j

-5 0 :::j

oc 1 Figure 12: Principal Component Analysis plot for basic call types TMP=Trumpet. RVM=Rumble. SQK=Squeak. SQL=Squeal. BL W=Blow. BRK=Bark. ROR=Roar PC I is represented primarily by spectral parameters. so clustering along PC l results primarily from variability in spectral structure. PC2 is represented primarily by temporal parametc-rs. so clustering along PC I results primarily from variability in temporal patterns.

Along PC l axis. the variability is explained by AM Rate Variation ( M22 ).

Overall Entropy (M26 ). and frequency parameters. namely Lower Frequency (M3 ).

Upper Frequency (M4). Bandwidth (M6). Median Frequency (Ml J ). Frequency of

Peak Overall Intensity (M20J. and FM Rate Variation (M24). The loading of AM Rate

V aria ti on (M22) and FM Rate V aria ti on (M24) is positive. so the value of these

61 parameters increase along the PCl axis. All other variables decrease along the PCl axis. Starting at the most positive end of the PCl axis, Rumbles have the lowest values for Lower Frequency (M3), Upper Frequency (M4), Bandwidth (M6), Median

Frequency (Mll), Frequency of Peak Overall Intensity (M20), and Overall Entropy

(M26), and the highest values for FM Rate Variation (M24) and AM Rate Variation

(M22). ·Moving from right to left along the PCl axis are Rumbles, Roars, Barks,

Squeaks, then Trumpets and Squeals with the highest values for Lower Frequency

(M3), Upper Frequency (M4), Bandwidth (M6), Median Frequency (Ml 1), Frequency of Peak Overall Intensity (M20), and Overall Entropy (M26) , and the lowest values for Freq1:1ency modulation (M24) and Amplitude modulation (M22).

Squeaks show the least within-call variation along the PCl axis, so this call type is the most stereotypic in its spectral structure. Trumpets show the greatest degree of variation along the PC 1 axis, so this call° type is highly variable in its spectral structure. These results are consistent with perceptual aural cues of lower and higher frequency calls, and of stereotypy in the Squeak call. It is difficult to perceive rate of change in modulation rates, so the results here provide differentiation that would otherwise be missed. Frequency jumps were perceived in the Squeal and changes tonality and noisiness were perceived in the Trumpet, so the results here for the variability in Entropy are consistent with aural cues for these call types.

Along the PC2 axis, the variability is explained primarily by AM Rate (M21 ),

FM Rate (M23), and temporal parameters, namely Duration (M5), Median Time (M7),

Temporal Concentration (M9). There is some contribution by Lower Frequency (M3),

62 Frequency of Peak Overall Intensity (M20), and Upsweep Fraction (M28), but the

contribution is small compared to the other parameters, so these were not included in

the interpretation of the plot. The loading of AM Rate (M2 l) and FM Rate (M23) is positive, so the value of these parameters increase along the PC2 axis. All other parameters decrease along the PC2 axis. Moving from top to bottom along the PC2 axis are Squeaks with the lowest values for the temporal parameters, followed by

Barks, Trumpets and Roars, then Rumbles, then Squeals with the highest values for temporal parameters.

Trumpets and Barks show the least within-call variation along the PC2 axis, so these call types are most stereotypic in their temporal pattern Squeaks and Barks are the shortest calls. Squeals and Rumbles include the longest calls and show the greatest degree of variation along the PC2 axis, so these call types are highly variable in their temporal pattern These results are consistent with perceptual aural cues of at least duration.

Five of the 6 parameters used as decision rule in the classification were contributors to the principal components. Only Up Sweep Fraction (M28) was not a contributor, and this parameter was used only to differentiate a small number of

Squeak calls. In the classification tree, Lower Frequency (M3), Upper Frequency

(M4 ), and FM Rate Variation (M24) separated most of the higher frequency calls

(Squeak, Squeal, Trumpet) from the lower frequency calls (Bark, Roar, Rumble), which is consistent with the separation along the PCI axis. In the classification tree,

Duration (M5) separated Barks from the other low frequency calls, Roars and

63 Rumbles. Roars and Rumbles were separated by Frequency of Peak Overall Intensity

(M20), which is supported by the separation along the PCl axis.

Analysis of Similarity (ANOSIM) The classification showed a high rate of successful classification, and the parameters that separated call types were consistent with the parameters that contributed to the variance in the Principal Component Analysis, but neither of these analyses determine if the difference among call types is significant. The null hypothesis states that there are no differences in acoustic parameters among the six basic call types (Bark, Roar, Rumble, Squeak, Squeal, Trumpet). A test statistic of differences among groups was computed using an Analysis of Similarity (ANOSJM).

Global and pair-wise ANOSIM tests of the scaled dataset were run to determine if there were significant differences in the acoustic parameters among the pre-defined call types and between pairs of call types. Squeak and Squeal and Roar and Rumble were sometimes challenging to differentiate using perceptual aural cues, and the grouping in the PCA showed overlap of these call types. ANOSJM tests were used to determine if these call types were significantly different or if they were gradations of a composite call type.

The global ANOSJM test showed significant differences among the call types

(R=0.432, p<0.001), as shown in Figure 13. The pair-wise ANOSJM tests with a

Bonferroni-corrected alpha (a=0.0033) indicated a significant difference between every call type, as shown in Table 9. Not surprisingly, the dissimilarity was greatest between the Rumble and the higher frequency calls, Squeak (R=0.807, p<0.001),

64 Trumpet (R=0.731, p<0.001) and Squeal (R=0.675, p<0.001). The call types that are challenging to distinguish by aural cues had mid-range R values, namely Squeal and

Squeak (R=0.425, p<0.001) and Roar and Rumble (R=0.417, p<0.001), so it appears that these call types are indeed distinct. The most similar pairs were Bark and Roar

(R=0.227, p<0.001), Bark and Rumble (R=0.278, p<0.001), and Squeal and Trumpet

(R=0.230, p<0.001), which is consistent with misclassification rates in the classification trees and the clustering overlap in the PCA plots. These calls were easy to distinguish by aural cues, so it may be that the acoustic structure is more similar across acoustic features that are difficult to perceive.

It is important to note that the group sizes were not balanced, so the statistical power is low, and the pairs with low R values may not be significantly different in a dataset with balanced group sizes.

R= 0432 P=

~ ' ' ' i~ ' ' ' ' ' ' ' :' ' ~~ I 1 .!.' I' I' ' ' :·' I ' ' ~~ lj ' ' ' ~ ' ' ' ! ' ' ' ' ' ~~ ' ' ' ' ' :' . ' ' ' ,' , ~ ~' ~ 0 ----'--- ~ ~ Between BRK ROR RUM SOK SQL TMP Figure 13: Global ANOSIM test for difference in parameters among the call types. TMP:::Trumpet, RUM==Rumble, SQK=Squeak, SQL=Squeal, BLW=Blow, BRK==Bark, ROR==Roar Plot shows that the difference between call types is larger than the difference within each call type.

65 Table 9: Summary of pair-wise ANOSIM showing significant difference between all call types

Call pair R value Significance Significant Difference (Bonferroni-corrected a=0.0033) BRKv.ROR 0.227 <0.001 Yes BRKv. RUM 0.278 <0.001 Yes BRK v. SQK 0.539 <0.001 Yes BRK v. SQL 0.448 <0.001 Yes BRKv. TMP 0.342 =0.001 Yes RORv.RUM 0.417 <0.001 Yes ROR v. SQK 0.685 <0.001 Yes ROR v. SQL 0.453 <0.001 Yes ROR v. TMP 0.468 <0.001 Yes RUMv.SQK 0.807 <0.001 Yes RUMv.SQL 0.675 <0.001 Yes RUMv. TMP 0.731 <0.001 Yes SQKv. SQL 0.390 <0.001 Yes SQKv. TMP 0.422 <0.001 Yes SQLv. TMP 0:230 <0.001 Yes

Descriptions of call types by acoustic parameters

Acoustic measurements using Osprey yielded comparative measurements of the basic call types (Bark, Roar, Rumble, Squeak, Squeal, Trumpet) and Blow. The parameters selected for describing the calls types were those that were shown to be most important in separating call types by the classification tree and PCA. Out of the

18 parameters that· were included in the analyses, 14 were shown to be important in differentiating call types. Four of the parameters (Bandwidth (M6), Median Frequency

(Ml 1), Median Time (M7), and Temporal Concentration (M9)) were highly correlated with other parameters and were not used by the classification tree, but were included in the final description for completeness. Table 10 lists median values for these 14 parameters across the basic call types for quick comparison of acoustic structure.

66 Table 10: Median values of acoustic parameters shown by the classification tree and PCA to differentiate basic call types

for the inner 90% of the signal energy~, rather than th ------r-.---- Bark Roar Rumble Squeak Squeal Trumpet

Lower Frequency (Hz) 75 124 26 495 1031 484 (M3) Upper Frequency (Hz) 1065 777 665 1830 2896 4620 (M4) Duration (sec) 0.5 1.5 2.3 0.2 1.9 0.5 (MS) Bandwidth (Hz) 991 657 583 1318 1873 4091 (M6) Median Time (sec) 0.22 0.66 0.98 0.10 0.75 0.24 (M7) Temporal Concentration (sec) 0.22. 0.74 1.21 0.10 1.07 0.27 (M9) Median Frequency (Hz) 408 322 165 912 1509 1305 (Mll) Frequency of Peak Overall Intensity (Hz) 172 237 133 775 1443 1098 (M20) AM Rate (Hz) 2.0 0.9 0.5 4.7 2.3 2.2 (M21) AM Rate Variation (Hz) 0.029 0.046 0.372 0.012 0.023 0.023 (M22) FM Rate (Hz) 2.3 0.8 0.5 4.9 1.5 2.7 (M23) FM Rate Variation (Hz) 0.006 0.012 0.093 0.003 0.006 0.006 (M24) Overall Entropy (Hz) 91 42 25 136 145 184 (M26) Upsweep Fraction (%) 46 50 44 34 47 49 (M28)

Table 11 provides the final acoustic ethogram for the basic call types (Bark,

Roar, Rumble, Squeak, Squeal, Trumpet) plus Blow. This ethogram describes calls both by aural cues and acoustic parameters. Values for Median Time (M7) and

Temporal Concentration (M9) in relation to Duration (M5) showed that energy was 67 distributed fairly evenly around the center of the call for all call types, so these values were not included in the final ethogram. What we can say instead is that the energy is distributed around the center of the call for all of the basic call types in the repertoire.

Upsweep Fraction (M28) was not included in the description because it differentiates only a small number of Squeaks.

All Osprey measurements are made on the upper 90% of the call energy, so range values may be greater for calls measured with signal processing tools that measure extreme values.

Table 11: Final ethogram for Bark, Roar, Rumble, Squeak, Squeal, Trumpet, and Blow Measurements are for the inner 90% of the signal energy rather than the entire signal. The spectrograms are not the same frequency range or duration. The images were captures to show the ranQ:e of the oarticular call rvoe. Call type and spectrogram Definition

Aural cue: Sounds like a short loud grunt, clearing throat. cape buffalo. chuff of cow. cheetah. Ends abruptly with no roll off. Noisy compared to rumble. Visual cue of caller: Mouth open. BARK Hz Measurements of90o/o signal energy Duration 0.1 - 1.3 sec (median 0.5)

8000 Lower Frequency 26 - 240 Hz (median 75) Upper Frequency 248 - 4963 Hz (median 1065) 6000 Bandwidth 129 - 4910 Hz (median 991)

4000 Median Frequency 101- 1182 Hz (median 408) Freq of Peak Intensity 37 - 840 Hz (median 172) 2000 AM Rate 0 - 21.5 Hz (median 2.0) AM Rate Variation 0- 0.186 Hz (median 0.0.29) 0.0 0.1 0.2 0.3 0.4 0.5 0.6 times FM Rate 0 - 24.8 Hz (median 2.3) FM Rate Varia ti on 0 - 0.046 Hz (median 0.006) Overall Entropy 17 - 468 (median 91)

68 Call type and spectrogram Definition

Aural cue: Sounds like a lion roar with changing pitch Pitch sounds like it changes because of amplitude modulation. Perceived as low frequency. Visual cue of' caller: Mouth wide open or widens during call. lower abdomen contracts.

Measurements of90% signal energy ROAR "' Duration 0.5 - 3. 7 sec (median 1.5) Lower Frequency 23 - 361 Hz (median 124) 4000 Upper Frequency 210 - 3763 Hz (median 777)

3000 Bandwidth 54 - 3650 Hz (median 657) Median Frequency 122 - 1214 Hz (median 322) 2000 Freq of Peak Intensity 40 - 1292 Hz (median 237) 1000 AM Rate 0.3 - 4.3 Hz (median 0.9)

00 0.5 1.0 1.5 2.0 AM Rate Variation 0.012 - 0.464 Hz hme s (median 0.046) FM Rate 0.3 - 6.8 Hz (median 0.8) FM Rate Variation 0.012- 0.232 Hz (median 0.012) Overall Entropy 11 - 208 (median 42 )

Aural cue: Sounds like a diesel engine. motorboat. distant helicopter. Pulsating with a flat pitch. Perceived as low frequency. Visual cue of caller: Mouth visibly open, forehead flutter.

Measurements of90% signal energv RUMBLE Duration 0.9 - 10.0 sec (median 2.3) GOO Lower Frequency 14 - J 20 Hz (median 26) 500 Upper Frequency 104 - 2498 Hz (median 665) ""' Bandwidth 86 - 2422 Hz (median 583) ""' i Median Frequency 26 - 590 Hz (median 165) ' ?00 ~ Freq of Peak Intensity 17 - 750 Hz (median 133) l 100 I AM Rate 0.1 - 3.6 Hz (median 0.5) AM Rate Variation 0.023 - 2.972 Hz 0.ti l 0 ,,,.. , 'l.t 3.0 J .~ '' I (median 0.372) FM Rate 0.1 - 2. 7 Hz (median 0.5) FM Rate Variation 0.023 - 0.929 Hz (median 0.093) Overall Entropy 4 - 133 Hz (median 25)

69 Call type and spectrogram Definition

Aural cue: Sounds like a wet finger across tight rubber surface. audible pedestrian crossing signal. baby alligator. gibbon call. High-pitched sound that falls in pitch. powerful initial sound. Visual cue of caller: Cheeks depressed. mouth open slightly or closed. tail sometimes erect. SQUEAK Measurements of 90% signal energv Duration 0. l - l.2 sec (median 0.2)

8000 Lower Frequency 194 - 883 Hz (median 495) Upper Frequency 797 - 3 768 Hz (median l 830) 6000 Bandwidth 301 - 3273 Hz (median 1318)

4000 Median Frequency 538 - 1905 Hz (median 912) Freq of Peak Intensity 537- 2024 Hz (median 775) 2000 AM Rate 0.8 - 44.9 Hz (median 4. 7) AM Rate Variation 0.003 - 0.026 Hz 0.0 0.1 0.2 0.3 0.4 11me. s (median 0.0.012) FM Rate l.l - 70.5 Hz (median 4.9) FM Rate Variation 0.003 - 0.032 Hz (median 0.003) Overall Entropy 58 - 294 Hz (median 136)

70 Call type and spectrogram Definition

Highly variable in strucrure and duration. Variations SQUEAL numbered in aural cues. Aural cue: ( l) Short utterance. Sounds like rubber shoes on a gym floor. release of air out of a balloon. windsrueld wiper on a dry window. a dog whine. (2) Train of repeated Squeal with short breaks between sounds. Sounds like squeals in staccato. intermittent release of air from a balloon. Each squeal sounds flat in pitch but they jump up and down. (3) Sounds like continual release of air from a balloon. fingers on a wet balloon. sliding pitch. sounds long. Visual cue of caller: Cheeks depressed. mouth open

8000 slightly or closed.

~ /;' 6000 Measurements of 90% signal energ'' · !• Duration 0.1 - 16.5 sec (median 1.9) """' Lower Frequency 226 - 1927 Hz (median 1031)

2000 Upper Frequency 851 - 8193 Hz (median 2896) Bandwidth 129 - 7816 Hz (median 1873) Median Frequency 575 - 2384 Hz (median 1509) Freq of Peak Intensity 422 - 2218 Hz (median 1443) AM Rate 0.1 - l 2.3 Hz (median 2.3) 0000 AM Rate Variation 0.006 - 0.087 Hz ~ /;' 6000 (median 0.023) l 0.1 - 23.6 (median 1.5) """' FM Rate FM Rate Variation 0.006- 0.041 2000 (median 0.006)

0.0 05 LO 2.0 Overall Entropy 35 - 479 (median 145 ) time!> "

71 Call type and spectrogram Definition

Aural cue: Sounds like a trumpet blast. Perceived as a higher frequency call. Visual cue of caller: maybe a lifting at base of trunk. trunk sometimes extended. tail sometimes erect sometimes (any behavior indicating extreme arousal).

TRUMPET Measurements of'90o/o signal energv Durati on 0. 1 - 2.4 sec (median 0.5 ) Lower Frequency 54 - 980 Hz (median 485) 8000 Upper Frequency 128 1 - 8926 Hz (median 4620) 6000 Bandwidth 1012 - 8742 Hz (median 4091)

4000 Median Frequency 146 - 4454 Hz (median 1305) Freq of Peak Intensity 129 - 353 1 Hz (median 1098) """' AM Rate 0.4 - 27.6 Hz (median 2.2)

0. 0 02 0.6 0.8 10 .., '-' " htne ~ " AM Rate Variation 0.006- 0.061 Hz (median 0.023) FM Rate 0.4 - 23.l Hz (median 2.7) FM Rate Variation 0.006 - 0.244 Hz (median 0.006) Overall Entropy 29 - 83 l Hz (median 184)

Aural cue: Forced exhalation through end of trunk. Not a vocalization. Visual cue of caller: air blast from trunk

Measurements of'90o/o signal energy

BLOW Duration 0.3 - 1.2 sec (median 0. 7) Lower Frequency 81 - 1512 Hz (median 372) "'°"" " Upper Frequency 1448 - 9318 Hz (median 3790) 15000 ' 'J Bandwidth 1217 - 8990 Hz (median 3273)

"'~10000 1 ! Median Frequency 531 - 4021 Hz (median 1156) 5000 - . - Freq of Peak Intensity 97 - 3295 Hz (median 894) - - - 0 - -· - AM Rate 0.8 - 3.6 Hz (median 1.4) QC 02 0.4 06 °' 1.0 1 2 AM Rate Variation 0.012 - 0.07 Hz (median 0.046) FM Rate 0.8 - 25.1 Hz (median 1.6) FM Rate Variation 0.012 - 0.046 Hz (median 0.012) Overall Entropy 102 - 875 Hz (median 286)

72 Nonlinear dynamics of vocal production These basic call types exhibited some nonlinear dynamics of vocal production,

which include subharmonics, deterministic chaos, bifurcations, biphonation, and frequency jumps (Mann et al., 2006). Subharmonics are sounds at frequencies other

than the fundamental frequency or integer harmonic of the fundamental that can be generated by coupled oscillators. Deterministic chaos comprises noisy signals that are

not random noise, but generated by chaotic process. Bifurcation is the rapid shifts between tonal harmonics and chaos. Biphonation is the generation of two independent frequencies simultaneously. Frequency jumps are abrupt up or down transitions between two frequencies or harmonics that are unpredictable (Mann et al., 2006). The

Trumpet exhibited clear evidence of bifurcation, as measured by the Overall Entropy

(M26) parameter and by visual inspection of Trumpet spectrograms. The Squeak and

Squeal also exhibited evidence of bifurcation as measured by the Overall Entropy

(M26) parameter, but it was not evident from visual inspection of the spectrogram.

The Roar and Bark appear to be noisier sounds, but the shift from linearity to non­ linearity is not evident. The Squeal exhibited clear frequency jumps when produced in a train of Squeals. The degree to which these non-linearities vary with perceived level of arousal is being investigated, but is not included here.

Comparison of low frequency calls to background noise

Power spectra of Paired Rumbles and noise for the three study sites were compared to investigate the potential for Rumbles being buried in background noise.

At the Oregon Zoo (Figure 14) the Rumble had multiple peaks below 200 Hz. The

73 noise floor started to rise below 300 Hz, with the peak below 1OOHz. From visual

inspection, the signal-to-noise ratio was less than 10 dB below 100 Hz, but increased

in the 100-300 Hz range. Given this ratio, it appears that signal energy below 100 Hz

could become inaudible due to noise in this environment. In the entire range of 15 Hz

to 20 kHz, noise declined more sharply from lower frequencies to approximately 2

kHz, then declined gradually or flattened. In one recording there was a slight increase

in noise in the 6-8 kHz range.

At the Elephant Nature Park in rural Thailand (Figure 15), the Rumble had

multiple peaks below 200 Hz. The noise floor started to rise sharply below 100 Hz,

with the peak below lOOHz. Noise had the highest power at low frequencies (below

approximately 30 Hz). From visual inspection, the signal-to-noise ratio was less than

10 dB below 100 Hz, but increased in the 100-300 Hz range. Given this ratio, it

appears that the signal energy below 100 Hz could become inaudible due to noise in

this environmentas well. Wind gusts were a factor at this site. In the entire range of 15

Hz to 20 kHz, noise declined sharply from low frequency to approximately 400 Hz,

then declined gradually across the range. In all selected recordings there was a notable

increase in noise in the 2-4 kHz and 6-11 kHz ranges. Based on aural cues at the time

of observation, noise in 6-11 kHz band may be sounds of insects or other arthropods.

At the Royal Elephant Kraal in urban Thailand (Figure 16), the Rumble had

multiple peaks below 200 Hz. The noise floor started to rise below 300 Hz, with the peak below lOOHz. From visual inspection, the signal-to-noise ratio was less than 10 dB below 200 Hz (except for recording at close proximity), and remained fairly low to

74 about 300 Hz. Given this ratio, it appears that signal energy below 200 or 300 Hz could become inaudible due to background noise in this environment as well. Wind gusts were a factor at this site. In the entire range of 15 Hz to 20 kHz, noise declined sharply from 15 Hz to approximately 400Hz, then declined gradually across the range.

In all selected recordings there was a notable increase in noise in the 6-11 kHz range.

With the exception of the increase in the 6-11 kHz range, the noise profile across the entire range of 14 Hz to 20 kHz looked more similar to the Oregon Zoo than to the

Elephant Nature Park, so the frequency band of the background noise maybe primarily a result of the degree of urbanization.

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

90 .,, .,, ~ ~ g 85 iii 80 ~ ~ i 75 g 80 fH" 70 fli V'v·A~ 75 65 70 60

55 ss~~-~-~-~-~-~-~-~-~-~ 1m 2m 300 4m 500 60J 700 ooo 900 1000 1100 60 m ~ ~ B D ~ 400 ~ 500 frequency, Hz frequency, Hz

Noise: 14Hzto1.2 kHz Noise: 14 Hz to 20 kHz

100 100.---.,.--~--.,.--~----~-.,.--~-~-~

95 00

90 BO .,, 85 g 80

j 75 fli 70

65 A •. ~r'V-w.,.,A 60 1 ~W\./W •V'v\r1;.

55 20 200 400 600 800 1000 1200 2oo:i 4000 6000 8000 10000 12000 14000 16000 18000 frequency, Hz frequency, Hz Figure 14: Power spectra of Rumble and background noise at the Oregon Zoo Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value. 75 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

fllr-~~r-~~~~~~~~~~~~~~~~ BOr-~~~~~~~~~~~~~~~~~~--,

75

'll "g BJ g 65 ~ ~ 55 ~" § 5_ 60 ~ 50 lg 45 55 40 50 /\ 35

~ 45 ~ 200 400 500 ~ ~ 1200 ~~40~-,60~~00~~,oo".:-~,20~~,40~~,60~-,-,eo".:--=-200".:--=-220._,_~ frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

lllr-~~~~~~~~~~~~~~r-~~~---. 80r-~~~r-~r-~~~~~~~~~~~~--,

75 70 70

66 60

50

40 35 ~" :J:J 100 200 m 400 500 500 100 soo oo:i 1000 1100 m = == ~1=1~1=1=1~ frequency. Hz frequsncy, Hz Figure 15: Power spectra of Rumble and background noise at Elephant Nature Park Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

76 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

85~~~~~~~~~~~~~~~~~~~~---,

80

75 70 .., .,, 70 g Ei5 . ~. f 65 j ro i § 60 i55 !li 55 50

45 50

40 45

35 40 ~ 200 400 600 800 1000 1200 100 200 300 400 500 600 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

10~~~~~~~~~~~~~~~~~~~~ 10~~~~~~~~~~~~~~~~~~~---,

60 60 .,, 50 55 .~ l c ~ ~ -! ro § i 40 ai ~ 45 .., 30

40

20 35 v~ 10 3J 100 200 30'.l 400 500 600 700 000 900 1000 1100 = ~~ ~1~1=1~1~1~ frequency, Hz frequency, Hz Figure 16: Power spectra of rumbles and background noise at the Royal Elephant Kraal Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

77 CHAPTER V: DISCUSSION The goals of contributing to the basic science of elephant communication were met by the results of this research. An acoustic repertoire of Asian elephants based on acoustic parameters was defined. A comparison of how the repertoire was used by groups of elephants and individuals was made. The manual methods of classification based on perceptual aural cues and visual inspection of spectrograms was compared to automated classification, and structural distinction among call types was validated.

The fact that a limited number of acoustic parameters differentiated call types suggests that these parameters could be used for detection of elephant calls. The next steps would be to determine if the parameter set could be reduced further, and to examine differentiation of elephant sounds from non-elephants, the data of which has already been collected and is ready to analyze. Finally, the descriptions of elephant sounds presented in this thesis serve as a basis for comparisons among captive and wild Asian elephants and between Asian and African elephants.

The final repertoire was defined by 6 basic call types (Bark, Roar, Rumble,

Bark, Squeal, Squeal, and Trumpet), 5 call combinations and modifications with these basic types as their constituent parts (Roar-Rumble, Squeal-Squeak, Squeak train,

Squeak-Bark, and Trumpet-Roar), and a sound that was produced frequently by many elephants, the Blow. Results suggest these call types are differentiated by 11 temporal and spectral parameters, and with future analyses this feature set may be able to be reduced further.

78 Given the high success rate of the classification tree using only 6 parameters as decision rules, it appears that this tree does have potential as a predictive model for classifying new calls into pre-defined call types. Consistent with aural cues was the separation of the higher frequency calls (Squeak, Squeal, and Trumpet) from the lower frequency calls (Bark, Roar, Rumble), and the confusion between Roars and Rumbles.

Separation of the call types with the Principal Component Analysis appeared to be in agreement with the separation by the classification tree. Finally, analysis of similarity tests showed significant difference among the 6 basic call types and between all pairs of these call types, so these call types are structurally distinct.

Data used in this study were collected in both urban and rural settings with many sources of noise and recorded with multiple recording systems, and results suggest that automated call detection is possible in varying recording situations, with various noise sources and using various recording systems.

Biological sources of variability within call types may include individuality of the caller, social context, level or arousal and state of motivation, and potentially the size of the animals and size of the head space. Some call types appear to be stereotypic in either temporal or spectral structure and other vary widely. The Squeak is a stereotypic call, the sound production of which includes manipulation of the cheek or lips. The Squeal is highly variable, but the sound production appears to be the same.

Little is known about sound production in elephants, but it may be that these sounds are produced by s.ome source other than the vibration of vocal cords alone.

79 Results of repertoire usage agreed with (Langbauer, 2000) in that there Was a

difference in calling pattern of males and females, with male elephants producing fewer call types than their female counterparts (Langbauer, 2000). The repertoire defined by this study suggests some differences between African elephants and Asian elephants. The most common vocalization of African elephants is the Rumble, which is highly variable and is used in multiple contexts and serves multiple functions (Soltis et al., 2005; Olson, 2004). The Rumble is produced by Asian elephants, but it was not the most common sound produced in this study. The Asian elephants in this study were captive or domesticated and were not wild, but the published calls of wild elephants by McKay (1973) suggest that at least some of the call types defined in this study are also produced by wild Asian elephants. Ecological differences in populations of African and Asian elephants include predator pressures, resource availability, group size, and human encroachment. Home range size may be a function of these differences, and communication modalities and distance may be a function of home range and group size. One could hypothesize that a reason for the Rumble being the primary call of African elephants is a greater need for long distance communication in order to maintain separation while remaining in contact for the purpose of resource utilization.

Future analyses for this dataset include completing the caller identification, investigating the communicative function of these sounds based on behavioral data already collected, investigating potential explanations for differences in call distribution among individuals, running unsupervised call classification as another

80 validation of established call types, doing a more rigorous analysis of signal-to-noise ratio at various frequency values, and running non-elephant sounds through a classification tree as a preliminary test for call detection potential.

81 REFERENCES

Baken, R. J. 1996. Clinical Measurement of Speech and Voice. San Diego: Singular Publishing Group, Inc. Blake, S. & Hedges, S. 2004. Sinking the Flagship: the Case of Forest Elephants in Asia and Africa. Conser1lation Biology, 18, 1191-1202. Blanc, J. 2008. Loxodonta africana. In: IUCN 2008. 2008 IUCN Red List of Threatened Species. . Downloaded on 11 May 2009. Campbell, G. S., Gisiner, R. C., Helweg, D. A. & Milette, L. L. 2002. Acoustic identification of female Steller sea lions (Eumetopiasjubatus). The Journal of the Acoustical Society of America, 111, 2920-2928. Charif, R. A., Ramey, R. R., Langbauer, W. R., Payne, K. B., Martin, R. B. & Brown, L. M. 2005. Spatial relationships and matrilineal kinship in African savanna elephant (Loxodonta africana) clans. Behavioral Ecology and Sociobiology, V57, 327-338. Choudhury, A., Lahiri Choudhury, D. K., Desai, A., Duckworth, J. W., Basa, P. S., Johnsingh, A. J. T., Fernando, P., Hedges, S., Gunawardena, M., Kurt, F., Karanth, U., Lister, A., Menon, V., Riddle, H., Rubel, A. & Wikramanayake, E. 2008. Elephas maximus. In: IUCN 2008. 2008 IUCN Red List of Threatened Species. . Downloaded on 11May2009. CITES. 2008. Appendices I, II, and III to Convention on International Trade in Endangered Species of Wild Fauna and Flora . Downloaded on 11 May 2009. Geneva. Clemins, P. J., Johnson, M. T., Leong, K. M. & Savage, A. 2005. Automatic classification and speaker identification of African elephant (Loxodonta africana) vocalizations. The Journal of the Acoustical Society ofAmerica, 117, 956-963. Cortopassi, K. A. 2006. Automated and Robust Measurement of Signal Features. Bioacoustics Research Program, Cornell Laboratory of Ornithology, Cornell University. Darden, S. K., Dabelsteen, T. & Pedersen, S. B. 2003. A potential tool for swift fox (Vulpes velox) conservation: Individuality of long-range barking sequences. Journal ofMammalogy, 84, 1417-1427. Deecke, V. B. & Janik, V. M. 2006. Automated categorization ofbioacoustic signals: A voiding perceptual pitfalls. The Journal of the Acoustical Society of America, 119, 645-653. Doluweera, G., Fernando, S., Gunasekera, C., Madanayake, A., Rossel, G., Seneviratne, L., Wijetunga, S. & Y apa, S. 2003. Infrasound - Can }t be more than a means of communication between Elephants? In: Symposium on Human-Elephant Relationships and Conflicts. Sri Lanka. Douglas-Hamilton, I. 1972. On the ecology and behaviour of the African elephant. In: Oriel College, pp. 257: Oxford.

82 Eisenberg, J. F., McKay, G. M. & Jainudeen, M. R. 1971. Reproductive behavior of the Asiatic elephant (Elephas maximus maximus L.). Behaviour, 38, 193-225. Fernando, P., Vidya, T. N. C., Payne, J., Stuewe, M., Davison, G., Alfred, R. J., Andau, P., Bosi, E., Kilbourn, A. & Melnick, D. J. 2003. DNA analysis indicates that Asian elephants are native to Borneo and are therefore a high priority for conservation. Plos Biology, 1, 110-115. Fischer, M. 2004. AZA Elephant TAG/SSP Regional Collection Plan. Saint Louis: Saint Louis Zoo. Fitch, W. T. 1997. Vocal tract length and formant frequency dispersion correlate with body size in rhesus macaques. The Journal of the Acoustical Society of America, 102, 1213-1222. Fitch, W. T. & Hauser, M. D. 2002. Unpacking "Honesty": Vertebrate Vocal Production and the Evolution of Acoustic Signals. Springer Handbook of Auditory Research, 16: Acoustic Communication, 65-137. Fleischer, R. C., Perry, E. A., Muralidharan, K., Stevens, E. E. & Wemmer, C. M. 2001. PHYLOGEOGRAPHY OF THE ASIAN ELEPHANT (ELEPHAS MAX/MUS) BASED ON MITOCHONDRIAL DNA. Evolution, 55, 1882- 1892. Fowler, M. E. & Mikota, S. K. 2006. Biology, Medicine, and Surgery of Elephants. Ames: Blackwell Publishing. Fristrup, K. M. & Watkins, W. A. 1993. Marine Animal Sound Classification. pp. 29: Woods Hole Oceanographic Institution Technical Report WHOI-94-13. Garstang, M. 2004. Long-distance, low-frequency elephant communication. Journal of Comparative Physiology A: Sensory, Neural, and Behavioral Physiology, 190, 791 - 805. Garstang, M., Larom, D., Raspet, R. & Lindeque, M. 1995. Atmospheric controls on elephant communication. The Journal of Experimental Biology, 198, 939-951. Heffner, R., Heffner, H. & Stichman, N. 1982. Role of the elephant pinna in sound localization. Animal Behaviour, 30, 628-630. Heffner, S. R. & Heffner, H. E. 1982. Hearing in the elephant (Elephas maximus): absolute sensitivity, frequency discrimination, and sound localization. Journal of Comparative and Physiological Psychology, 96, 926-44. IUCN. 2008. 2008 IUCN Red List of Threatened Species. . Downloaded on 11 May 2009. Janik, V. M., Sayigh, L. S. & Wells, R. S. 2006. Signature whistle shape conveys identity information to bottlenose dolphins. Proc Natl Acad Sci US A, 103, 8293-7. Kemf, E. & Santiapillai, C. 2000. Asian Elephants in the Wild. In: WWF Species Status Report (Ed. by Wilson, A.). Gland, Switzerland: WWF-World Wildlife Fund for Nature. Langbauer, J., William R., Payne, K. B., Charif, R. A., Rapaport, L. & Osborn, F. 1991. African elephants respond to distant playbacks of low-frequency conspecific calls. The Journal ofExperimental Biology, 157, 35-46.

83 Langbauer, J., William R., Payne, K. B., Charif, R. A. & Thomas, E. M. 1989. Responses of captive African elephants to playback of low-frequency calls. Canadian Journal of Zoology, 67, 2604-2607. Langbauer, W.R., Jr. 2000. Elephant Communication. Zoo Biology, 19, 425-445. Larom, D., Garstang, M., Lindeque, M., Raspet, R., Zunckel, M., Hong, Y., Brassel, K., O'Beime, S. & Sokolic, F. 1997. Meteorology and elephant infrasound at Etosha National Park, Namibia. J. Acoust. Soc. Am., 101, 1710-1717. Mann, D. A., O' Shea, T. J. & Nowacek, D. P. 2006. Nonlinear Dynamics in Manatee Vocalizations. Marine Mammal Science, 22, 548-555. Marter, K., Quine, D. & Marler, P. 1977. Sound transmission and its significance for animal communication II: tropical forest habits. Behavioral Ecology and Sociobiology, 2, 291-302. McComb, K. 1996. Studying Vocal Communication in Elephants. In: Studying Elephants (Ed. by Kangwana, K.), pp. 112-119. Nairobi, Kenya: African Wildlife Foundation (AWF). McComb, K., Baker, D.R. L., Moss, C. & Sayialel, S. 2002. Long-distance communication of social identity in African elephants. Animal Behaviour, 65, 317-329. McComb, K., Moss, C., Sayialel, S. & Baker, L. 2000. Unusually extensive networks of vocal recognition in African elephants. Animal Behaviour, 59, 1103-9. McComb, K., Reby, D., Baker, L., Moss, C. & Sayialel, S. 2003. Long-distance communication of acoustic cues to social identity in African elephants. Animal Behaviour, 66, 195-195. McKay, G. M. 1973. Behavior and ecology of the Asiatic elephant in southeastern Ceylon. In: Smithsonian Contributions to Zoology, pp. 1-113. Washington, DC: Smithsonian Institution Press. Mellinger, D. K. & Bradbury, J. W. 2007. Acoustic Measurement of Marine Mammal Sounds in Noisy Environments. In: Second International Conference on Underwater Acoustic Measurements: Technologies and Results. Heraklion, Greece. Moss, C. J. 1983. Oestrous behaviour and female choice in the African elephant. Behaviour, 86, 167-196. O'Connell-Rodwell, C., Wood, J., Rodwell, T., Puria, S., Partan, S., Keefe, R., Shriver, D., Amason, B. & Hart, L. 2006. Wild elephant ( Loxodonta africana ) breeding herds respond to artificially transmitted seismic stimuli. Behavioral Ecology and Sociobiology, 59, 842-850. O'Connell-Rodwell, C. E., Wood, J. D., Kinzley, C., Rodwell, T. C., Poole, J. H. & Puria, S. 2007. Wild African elephants (Loxodonta africana) discriminate between familiar and unfamiliar conspecific seismic alarm calls. The Journal of the Acoustical Society ofAmerica, 122, 823-830. Olson, D. 2004. Elephant Husbandry Resource Guide. pp. 280: International Elephant Foundation.

84 Payne, K. B., Jr., W.R. L. & Thomas, E. M. 1986. fufrasonic calls of the Asian elephant (Elephas maximus). Behavioral Ecology and Sociobiology, 18, 297- 301. Payne, K. B., Thompson, M. & Kramer, L. 2003. Elephant calling patterns as indicators of group size and composition: The basis for an acoustic monitoring system. African Journal of Ecology, 41, 99-107. Poole, J. H. 1987. Rutting behavior in African elephants: the phenomenon of musth. Behaviour, 102, 283-316. Poole, J. H. 1989. Announcing intent: the aggressive state of musth in African elephants. Animal Behaviour, 37, 140-152. Poole, J. H. 1999. Signals and Assessment in African Elephants: Evidence from playback experiments. Animal Behaviour, 58, 185-193. Poole, J. H. in prep. Elephant vocal communication. In: The Amboseli Elephants: A long-term perspective on a long-lived mammal. Chicago: University of Chicago Press. Poole, J. H., Payne, K. B., Jr., W.R. L. & Moss, C. J. 1988. The social contexts of some very low frequency calls of African elephants. Behavioral Ecology and Sociobiology, 22, 385-392. . Reby, D., Joachim, J., Lauga, L., Lek, L. & Auglagnier, S. 1998. Individuality in the groans of fallow deer (Dama dama) bucks. Journal of Zoology (London), 245; 79-84. Rendall, D., Owren, M. J. & Rodman, P. S. 1998. The role of vocal tract filtering in identity cueing in rhesus monkey (Macaca mulatta) vocalizations. The Journal of the Acoustical Society ofAmerica, 103, 602-614. Riddle, H. S., Rasmussen, B. & Schmitt, D. L. 2003. Are captive elephants important to conservation? Gajah, 22, 57-60. Roca, A. L., Georgiadis, N., Pecan-Slattery, J. & O'Brien, S. J. 2001. Genetic Evidence for Two Species of Elephant in Africa. Science, 293, 1473-1477. Sanvito, S., Galimberti, F. & Miller, E. H. 2007. Vocal signalling of male southern elephant seals is honest but imprecise. Animal Behaviour, 73, 287-299. Schulte, B. & Rasmussen, L. 1999. Signal-receiver interplay in the communication of male condition by Asian elephants. Animal Behaviour, 51, 1265-1274. SEVP. 2008. Savanna Elephant Vocalization Project . Shoshani, J. 2000. Elephants: majestic creatures of the wild. New York: Checkmark Books. Shoshani, J. & Eisenberg, J. F. 1982. Elephas maximus. Mammalian Species, 1-8. Soltis, J., Leong, K. & Savage, A. 2005. African elephant vocal communication II: rumble variation reflects the individual identity and emotional state of callers. Animal Behaviour, 70, 589-599. Stoeger~Horwath, A. S., Stoeger, S., Schwammer, H. M. & Kratochvil, H. 2007. Call repertoire of infant African elephants: First insights into the early vocal ontogeny. The Journal of the Acoustical Society of America, 121, 3922-3931.

85 Sukumar, R. 2003. The living elephants: evolutionary ecology, behavior, and conservation. New York: Oxford University Press. Sukumar, R. 2006. A brief review of the status, distribution and biology of wild Asian elephants Elephas maximus. International Zoo Yearbook, 40, 1-8. Vidya, T. N. C. & Sukumar, R. 2005. Social organization of the Asian elephant (Elephas maximus) in southern India inferred from microsatellite DNA. Journal ofEthology, 23, 205-210. von Muggenthaler, E. 1992. Infrasound from the okapi. In: American Association for the Advancement of Science (A.A.A.S) 158th conference. von Muggenthaler, E. 2000. Infrasonic and low-frequency vocalizations from Siberian and Bengal tigers. The Journal of the Acoustical Society ofAmerica, 108, 2541. von Muggenthaler, E., Baes, C., Hill, D., Fulk, R. & Lee, A. 2001. Infrasound and low frequency vocalizations from the giraffe; Helmholtz resonance in biology. In: Acoustical Society ofAmerica, regional conference. South Carolina: Proceedings of Riverbanks Consortium. von Muggenthaler, E., Reinhart, P., Lympany, B. & Craft, R. B. 2003. Songlike vocalizations from the Sumatran Rhinoceros (Dicerorhinus sumatrensis). Acoustics Research Letters Online (ARLO ), Acoustical Society of America, 4, 83-88.

86 APPENDIX A: DISTRIBUTION PLOTS OF PARAMETERS

IDSTOGRAMS AND BOXPLOTS OF ORIGINAL DATA

All of the histograms and boxplots below include. outliers. The summary statistics exclude outliers removed. Plots are included only for the 18 parameters that were used in the statistical analysis.

The figure panels are as follows: A) Histogram of parameter for all call types combined. B) Histogram of parameter for each call type. C) Summary statistics for all call types combined. D) Box plot of parameter for all call types combined. E) Boxplot of parameter for each call type.

Abbreviations for the call types are as follow: TMP (Trumpet), RUM (Rumble), SQK (Squeak), SQL (Squeal), BLW (Blow), BRK (Bark), ROR (Roar).

A. B. ·c.

LowerFreqM3 "".. Min. : 14 .13 lstQu.: 188.42 20 40" Median : 495.26 21) Mean : 532.77 ~ 15 0 3rd Qu.: 721.36 l- o '"' 1 Max. : 1927. 22 i t-0 " ~ ll. " " "

100 500 1000 15&D 2000 dta.6$LowerFreqM3 .." " "

0 SOO 100tl 1SOG :2000 o soe 1000 1soo 2ooe d!a6$LowerF"'qM3 D. E. 2000

1500 ·--i .., ;; ' :~-] "' 1000 ' I• c . ,,.. " ~., ;;; !: • 0 . -' 0 L~-·- -~!"-- ' 500 [..-: 0 rj' 4=J ' __ @__ 0 __ J._' __ ' D' 0 ' ' 1000 is~ ~ c.tJ a::; "' Lowt'f'FreqM3 "" BLW BRK ROR RUM SQK SQL TMP cau Figure 17: Histograms and boxplots of Lower Frequency (M3) TMP and SQL more variable, other call types more stereotypic.

87 A. B. c.

UpperFreqM4 60 Min. : 103.6 1st Qu. :1341.8 40 Median :2261. 0 " 20 Mean :2780.7 3rd Qu.:3434.5 " ...OIR.• :'U :>~soli<:' "Y~:lE:;H: u:sai:;y;: · · Max. : 9318. 5 ~ 60 ! 1~ l­ .. 40 ~ o t' ! " ..e 20 a." LL1_.1_. BlW BRK ~i!' ROft 60 .<000 JOOD 5000 ""'' dta 6Sl/ppe!'FreqM4 ::J_L..1-11. L a 2000 6COO 10000 0 21JOO 6000 10000 dta.6SUpperFreqM4

D. E.

8000 ' ' ,...... , ' i I ":::. 6000 .,O' rLJi ~ . .-....,.~. > ,;.-, ,. ~., I ! 4000 • r 1 • " L,.._J 1 I • 2000 y Q __ , ·····~·-· c:p __ !_ __ 2CIOO 400Q ~ ~__ .!.~- ~ UpperFieqM4 "'"" Bl.W BRK ROR RUM SQ)( SOL TMP cau Figure 18: Histograms and boxplots of Upper Frequency (M4) ROR and RUM have narrow range. TMP, SQL, and BLW highly variable, which may be partially a function of signal strength (distance to caller).

88 A. B. c.

;.{H:t'if;:~'ii'MP.?:t>:\q;.~ DurationM5 100 Min. : 0.067 80 1st Qu.: 0.331 60 Median : 0.737 H 40 "" 11111 20 Mean : 1. 292 3rd Qu. : 1. 5 7 9 0 ) SOK ,,., . ., ::: :i;L RIJM\''::::v: Max. : 14. 820 1U 100 0 f- 80 0 60 c.. 40 ~ 20 l:JL I Q_" L f BlW LILBRK :1 ROR 100 80 60 1~ " 4l) a:a.6$0ISabonM5 I 20 L L L 5 10 15 5 10 1'5 dta6$DurationM5 D. E.

15

:::< 10 "'c J 0 _J ~ n, __ .. ~ :

e [~]__ ., __ Q :.l .. " ~ ~~ ~ .J..,. i::;;:Q. DuJauonM5 BLW BRK ROR RUM SOK SOL TMP Call Figure 19: Histograms and boxplots of Duration (MS) ROR and RUM overlap in perception. SQL variable in duration. SQK, BRK, and TMP stereotypic.

89 A. B. c.

BandwidthM6 6-0 Min. : 53.83 50 1st Qu.: 904.39 40 30 Median :1682.08 20 Mean :2247.91 1D i ~ 3rd Qu. :2802.61 0 ,,,, ----- y,_ ----~·~•'-'" _,, , , ~~" · ·, , I Max. : 8990-11 1il 6-0 0 50 ! " t- " 0 j 1:.. ~ " "-" l Btw 60 50 40 ""' 30 dta,6SBa.'IC1Wialh~ 2ll 1-0 0 0 2000 6-000 0 2000 6-000 dta. 6$8andwid1hM6 D. E.

8000 '

; "'::> 6-000 r-i ~ l3 ~ J_ n ~ 4000 ,; I! ~ .. ! • ! c al --~-- y~ i I --r- ' 2000 ! I -~--- ,---:--, 8 ' ' 200<) G.J ,--•:::i .... J.' •• _ [tJ ~~~T~-'. BandwidthM6'''" "" """' BtW BRK ROR RUM SQl< SQL TMP can Figure 20: Histograms and boxplots of Bandwidth (M6) RUM mostly low frequency (as expected). BLW and TMP have very high bandwidth and variability, which may partially be due to distance to source. Squeal has relatively high bandwidth and very high variability. SQK has relatively wide bandwidth, but it is stereotypic.

90 A. B.· c.

C'' TMe:!cA'''!l MedianTimeM7 100 Min. 0.0210 BO 1st Qu.: 0.1340 60 Median : 0.3245 40 Mean : 0.5950 20 3rd Qu.: 0.6773 .. 0 FITL RUM SOK:;::_: '.;o;Jh <·/"'l;-.rw:·.··, Max. :10.5830 10 ~ 80 ! 1-- ~ .. 0 60 j c: 40 .,"'~ 20 ll. ,. 1L1t. " fr Blil'VL: '.'.'}; BRK !:i JIDR 100 80 60 40 d:a.6SMedlaniimeM7 I 20 0 LILI~10 10 dta.6$MedianTimeM7 D. E.

12

10

:i 8 ...... , .. . ~ E ~ 1;· r •M~-a·~~ ~.. • '5., • " L~., ::;; 4 "''i''"- --l- _J __ e 0 __ ... __ ~ g::: ct 0 Si~.? .1. ~ i:iik~ MectanT1meM7 " " BLW BRK ROR RUM SQK SQL IMP Call Figure 21: Histograms and boxplots of Median Time (M7)

91 A. B. c.

TimeConcentM9 100 ™"".''''' Min. : 0. 0000 80 1st Qu~: 0.1340 60 Median : 0.3370 40 20 Mean : 0.7946 0 3rd Qu. : 0. 8590 •L RUMii>i"''' : SOK SOL . Max. :13.8450 100 2c >- 80 0 60 1.. i I r c.. 40 I ..e kn_ 20 Q. IL i BLW LBRK '.; ROR 100 60 60 40 ~ 6STtmeConcer.1M9 I 20 L L L 10 15 10 15 dta. 65TimeConcentM9 D. E.

"' 10 ~ ~ j[ ! c s ~· ~---~--"" <>->-»->.> 0 ~ 0 ~ .~ 5 --~-- ~L >- --~--

-···y .. •· ' [;] --~-­ " ~ ~~·;".J ~ .4- LJ ~ TimeConcentM9 BLW 8RK ROR RUM SQK SQL TMP Gall Figure 22: Histograms and boxplots of Temporal Concentration (M9) SQL highly variable. RUM more variable than others.

92 A. B. c.

%',:·:: ':TMP TimeAsymmMlO 100 Min. :0.39400 80 1st Qu. :0.00000 6(} Median :0.00000 40 Mean :0.0006586 20 l_ 0 3rd Qu. :0.00100

! 60 ~ I- 0 ,070600 " c 'liir= l ..e "-.. \&:\fr i BlW H''.lE::.\BRK ROR 100 - 60 60 .. ea a..~ 40 ala.€STimeAs)'mrr.M10 20 1-l~ll_I I i I I I I I I I I I I I I I I ( l I -0.4 O.Q 0.20.40.60.B -0.4 0.00.20.40.60.8 dta.6$TimeAsymmM10 D. E.

0.6

0 0.4 ~ E !! .. ~ I . 0 ~ , 02 ~ .: --·~ ~ ! 0 . E o.o -+- ~ ...... -t- i + 8 > i= -+ + -0.2

.... ~:J.l! 0.2 ... '·' TimeAsymmM10 " -0.4 SlW SRK ROR RUM SOK SOL TMP Caff Figure 23: Histograms and boxplots of Temporal Asymmetry (MlO) Stereotypic with minimal variability.

93 A. B. c.

'IEi'~; MedianFreqMll Min. : 26. 24 60 1st Qu.: 663.81 40 Median :1046.76 20 Mean :1105.67 " 3rd Qu. :1489.14 ;~;;.·· Max. :4454.33 ~ 60 ! l- ,, " o i c 40 Q_" .."re 20 Q_

60

40 1000 2!:-00 ::000 dl:t 6.SMediatlFreqM 11 20 LIL 0 1000 3000 0 1000 3000 dta6$MedianFreqM11 D. E.

4000 ~------1 '

~ 3000 0 :t------lLJ l · ---·-·-··--····L.: i " : ~ : . 3 . 0 . i • .. -r .... 0 8 [J 1000 i-j """ '""' '"' 1000 . Ll MediarFreqM11 . cb --g--- s --L-. c!J ~ ~;:~

BtW BRK ROR RUM SOK SOL ThlP Call Figure 24: Histograms and boxplots of Median Frequency (Mll)

94 A. B. c.

FreqAsymmM14 Min. :-0.63400 6() 1st Qu.:-.00600 40 Median :0.02800 20 Mean :0.05637 3rd'Qu. :0.12400 0 ..... i')'.t!ii:\•iRVM ·-· Max. : 0. 66600 ~ ! JO f- >; [ 60 0 40 ! 1: !T." .. 2D ~ _l_ I Cl. aL " r.; .. :.. •.: BLW BRK 60 40 " eta 6$F'"&!'.;AsyrnmM14 2D _lj ~,___._ I I ! I I I I -0.5 0.0 05 -0.S 0.0 0.5 dta. 6$FreqAsymrnM 14 D. E.

g 0.5 4 --:-- , ' ' ' ... 0 ' ··-1·· ' ' --?-- :1 --I~ - ·- E n E (j-1; GJ .~ . -1·----· . «<•<>:•!!<>< o.n GJ L!J . 4::J ~ ~ I I I .f0- --~-- . 8 --6-- ~ ' 0 __ ._ __ u. ' ' . ··:.

-0.5 -G.5 00 "eqAsymmM14 " ' BLW BRK ROR RUM SQK SOL TMP Cal Figure 25: Histograms and boxplots of Frequency Asymmetry (M14)

95 A. B. c.

,,,lifV !TMF':0',S'.1;1 PkOverallRelTM18 25 Min. : 0.00 20 1st Qu.:23.50 15 Median : 41. 50 10 5 Mean : 41. 89 3rd Qu. : 5 7 . 7 3 Max. : 99. 63 25 ! 10 ~ 20 0 15 j c., 10 ..l: 5 Cl.

25 20 15 tll.6$?1(QveralR~18 10 5

0 20 40 60 80 100 0 20 40 60 so 100 dta,6$Pk0vera11RelTM18 D. E.

100 --;-- --i-- ' ' BO ' ' :ii"' ' !:: 60 ' Q:;.. ~ ~., ! • > 40 " 0 • ""Cl. ~ LJ 20 ' ' ,. ' ' D~ , ...... --i--' PkOverallReffM18 " 0 Bl'IV 8RK ROR RUM SOK SOL 1MP Gaff Figure 26: Histograms and boxplots of Relative Time of Peak Overall Intensity (M18) High variability within call type, but similar among call types.

96 A. B. c.

PkOverallFM20 00 Min. : 17. 50 60 1st Qu.: 481.71 40 Median : 861. 33 20 Mean : 931.14 0 3rd Qu. :1346.23 K sot'"· Max. :3531.45 60 % ~ ~ $ I- 60 0 i c.. 40 .r" ~ 2\) Q." 0 ao 60 dta.6$PkOYeralFM2o =~ ill LIL 0 1000 2000 3000 0 1000 2000 3000 d1a.6SPkOVeraHFM20 D. E.

3000

§ ' !!,, 2000 ' <'l ~ ~ § ' Q. 1000 ~ L:J --1-· lTJ e Q __ i.._. 1000 ""' 3000 GJ w ~~-~~ PkO\'el'allFM20 BtW BRK ROR RUM SQK SQL TMP Galt Figure 27: Histograms and boxplots of Frequency of Peak Overall Intensity (M20)

97 A. B. c.

~~' AMRateM21 100 Min. : 0.000 80 1st Qu.: 1.118 Median : 2.207 " Mean : 3.053 ~~L_ 3rd Qu. : 3 . 7 4 5 Max. :44.939

! r-~ ~JO 0 c., ~" ..~ CL LfLILt~BLW BRK ROl'Ui: -

lO dla 6SAMRa!et.121 ~lLLL 010203040 010203040 dla.6$AMRateM21 D. E.

4V

30 11·1 ~ ~ " f ' ~ ;;; w,---.--»-~ ll'. 21l ~ __ iL 10 . --V-- --'-- _J __ .L . __ _. __ C!J .. ~ r.::::P c::b ~ ' ~~~ L-'--...i " AMRa!eM21" " Bt.W BRK ROR RUM SOK SOL TMP Call Figure 28: Histograms and boxplots of AM Rate (M21)

98 A. B. c.

AMRateVarM22 100 Min. :0.00000 80 1st Qu. :0.01200 6-0 Median :0.02300 40 Mean :0.05642 20 0 3rd Qu. :0.04600 cRUM t'. SOK k' <::ru Max. : 2. 97200 ~ ~ I- ~ Ci j~ c .,"'~ 0. I~L[L/L![ .. · . 61.W ) . · SRI(;;'.::!:;:>:: ROR llilllll 100 80 60 dd.6SAMRateVa11-R2 40 20 0 LILIL 2 dta.6SAMRateVarM22 D. E.

3.0

2.5

~ 2.tl ~ f ~ ~ il~i . .,, ><·<' ~ 1.5 ii Cl'. " ~ i ~ 1.0

OS 0 ' 1 0 ' B 1.5 3.0 ~ ~ r.:= ~· _. __ '·' " " " 00 A.\1RafeVarM22 i --- --Bl.W BRK RQR RUM SOK SOL Th'IP Call Figure 29: Histograms and boxplots of AM Rate Variation (M22)

99 A. B. c.

i~~ea~71:~~~"~ FMRateM23 100 Min. : 0.0000 80 lstQu.: 0.9653 60 Median : 2.2160 40 Mean : 3.7613 20 L 3rd Qu.: 4.4170 Max. :70.4720

~ ~ ~ 4~ f- 0 i c., .,~ "- LiBtW BRJ'

D. E.

60

"'N :11 ; ' i 40 ~ :J· -- +--0 ~. C! ~ 'LJ : a: 20 i 0 0 0 • 0 --~-- __; __ 0 ~ --1'-- -l- ~ ·---~---' .. ~ ci;:J 0 ± Ml~ ~ c:t::i "' FMRateM23" BLW BRK ROR RlJM SOK SQL TMP Call Figure 30: Histograms and boxplots of FM Rate (M23)

100 A. B. c.

Hi TMP FMRateVarM24 100 Min. :0.00000 so 1st Qu.:0.00600 60 Median :0.00600 40 Mean :0.02094 20 .. 0 3rd Qu. :0.01200 ,.L RUM ll! "rn Max. :0.92900 100 .... ~ i' t- 60 ~ 0 60 ~ ~ ~ c 40 ..e Q) 20 a. ~ 0 "' i/ Bl..W LBRK .. LROR 100 ' - so ... ., .. ... ,, 60 ~.GSFMR.i.!e'lart'.424" 40 20 0 L L L 0.0 02 0.4 0.6 0.8 1.0 0.0 0.2 04 0.6 0.8 1 0 dta .6$FMRateVarM24 D. E.

¢

o.s 0

~" 0.6 Ill ~ .; . - .• ' '0 ·• ~ < ~ 0.4 0 ::;, t: LL ,, ··-r-- . 0.2 ., <> ii --T-- I • ; 8 l.._.,.. __I D.D {i.2 o.e c~:J _._ _j_ '·' " 0.0 ~ --1-- FMR8teVarM24 -·-- BlW BRK ROR RUM SOK SOL IMP Call Figure 31: Histograms and boxplots of FM Rate Variation (M24)

101 A. B. c.

b'ff'.z".~1MF:'n ·. EntropyM26 80 Min. : 3. 968 &ll 1st Qu.: 79.487 4l) Median :136.609

20 Mean :159.961 3rd Qu. :194.310 ,, Max. :1000. 792 ~ 80 ! I­ 60 " C; J:0 c 40 "'!,> 2{) &'. _, ..,,, I - "' tia.6$Eri!ropyt.126 :j _L.__ I 1 I 1 I I I I I I I I I I I I l I l 0 200400600 800 rL0 200400600 800 dta 6$EntropyM26 D. E.

1000

' ' ' .. ; --,-- ~ ·-----~-1·~·>~·;

102 A. B. c.

,ii{1!ki~ll'~Pi :v,::,ij UpswpFracM28 Min. : 6 .196 1st Qu.: 40.340 Median : 46.450 Mean : 44.931 3rd Qu.: 51.132 RL ;>r,,:;SQL Max. :100.000 ]! 0 ! f- " 0 . c j "' Q_""' _lJ~

dta..65UoswiiFracM28

0 ' ' .~ lLl_tJ0 2Q 4() 60 80 100 20 40 60 ..100 1 o ao ctta,6$UpswpFracM28 D. E.

100

SG

a:! "' --7-- --T-- ; Di''···,,,.,... ! ~ --r- ; ! : ~ --r- u.. ' ~ C, --:-- .:•<> -----r------1 .• ------r·<. .JN~ :;.;.< [~] 'ii: ~ [5 ' I ' 0, 40 C;J r-i ; 1 : ·1 .....~l lfJ => I • I ' ' ! ! -··i-- L-,-1 20 -+- "' .. " UpswpFracM28 " "' BLW BRK ROR RUM SQK SQL TMP Call Figure 33: Histograms and boxplots ofUpsweep Fraction (M28)

103 Q-Q PLOTS OF SCALED DATA

In a Q-Q plot, the x-axis is the expected value for a normal distribution and the x-axis is observed. The data are ranked and the quartile is calculated. If data are distributed normally then the shape will be a straight line where y=x.

-3 -2 ·1 G 1 2 J

"J ;,>Cc' ..,,. ~

M 2 ::> ~ f--1 ~ ..,.,,,,---~ "I . l ' ../l.L/ ..,.· ~ , <' ... :BRK:';·!'.<'.".'"dli".'J'.ROR·::;·.:-->:Tc• !'.'.'hlUJl,F 0 ' l i \:: U> J ~ ~

., ., .,~· .,.~~· ~··

·2 -3 -2 -1 G ,1 2 3 .J -2 -1 0 1 2 3

·3: -2 -1 0 1 2 3 b'::. •-:t::.S:OJ< !) ·;"L'..·•<''.•JMFl <.d '.· . ' (" /,•' 1 / i ~ .~··1 .. / ,, -1 ~ i' I ~ :oo<"·~ ~+(t~~ b~/-~f~'.:'::1u1u~·,._, ,,,,.,'T i 3 ~ ~ '

-1 -1 ,_,.!' <00___.. .~

·2 -3 -2 -1 0 1 2 -3 -2 -~ I) l 2 3 (JlOrm qnorm Figure 35: Q-Q plots of Upper Frequency (M4)

104 .3 -2 -1 I) 1 2 3

"' "g .!! • ~ I ! 1·.:u.,:;>E!f\1(.;;.: i .:; . ~ ~

/ ,.,-«'' .~' .,/

·2 -3 ·2 -1 c- 1 2 3 .3 -2 -1 0 1 2 :J C/l(lml qnorm Figure 36: Q-Q plots of Duration (M5)

.3 ·2 ·1 0 1 2 3 ·;·· ;:;·-:;;SIJKi "· ... ; TM"ff';

<.; ·~

"' ' /' i I . ,.,..,,.,, ~ I ./'/ -I :. Ill' ! +i :;:;:i'!iF!~ •· ''.f\OJ<:.': · '"l' < :c

,ff

-1 ,,_.,,/ ,,__..-8 ,,,~-·

-2 -3 -2 -1 0 1 2 '3 .J ·2 -1 0 1 2 3 qnorm qnorm Figure 37: Q-Q plots of Bandwidth (M6)

--'-~~~~~~~~~~ -3 -2 -1 0 1 2 3

10 10

~

~ i j ,,•1 / ~ 6 ,, QQ~ ')'.~"P<;. ~ :~'.; J~.f1C';'' :;<'>--;- I sn i ! ~ 1Q i 4 "' ii '

,/' ,,___.,.,,-- ,_..../ ~~'<.~'"'

·2 -J -2 -1 0 1 2 J -J -2 -1 0 1 2 J qnorm qnorm Figure 38: Q-Q plots of Median Time (M7)

105 .3 ·2 ·1 0 1 2 l

·::-£-,./'.~ :S.-Qt<.e;:;_ ·!']:-·.:,: :·0.··JMP " ,. / ~ i 5 ii § ,____.,."f-o u

,__/ ,oo_/j~ ,-"' "~

·2 .) -2 -1 0 1 2 3 .3 -2 -1 i) 1 2 J qnorm qnoon Figure 39: Q-Q plots of Temporal Concentration (M9)

..J -2 -1 0 1 2 3 1 20 ii>YiU'-.SOW ·., .·sot ;. :) ..:-:it:\:: .!::ndf!:' Y·'·,.

"'15 15 10 ___.. ,,--I"' 0 ,. ~ 0 §.. .. ~( <'> -10 i 5 ,, 1 ../ ':i :.:.u i:;imKin1r .• :hF"'··· -,.RCIF: · ·:: .i.• : ·:::m1W•:' ·· ~ ______E 20 "' v ~ 15 ~ r ~ 10

~

,,"'~~<;?

-2 -3 -2 -1 0 1 2 3 .J -2 ·1 -0 1 2 qnonn qnoon Figure 40: Q-Q plots of Temporal Asymmetry (MlO)

..J ·2 ·I O 1 2 3 ,,, '"'" :: '"'' ;;:,::so , ;~ l '.,_TMP~',';_ ··" .:'"' i •"

-~·: .'!{' ,__.,,,-· " -~ I ./·1,/ { LL •" 'O Ii •. ,.·<:i!lRttc>• · :·Roo: •. · .]7j;'. ·f'RDi.I ~ a

__,.../ ·•· I ___,,t••P '"°"n.:• (-'>-~ 6~· .__...,.,.·

·2 -3 ·2 ·1 0 1 2 .3 -2 -1 0 1 2 3 qnorm qnorm Figure 41: Q-Q plots of Median Frequency (Mll)

106 .J -2 ·1 0 l 2 3

,/'<>c ff' : ;t

" .,. /l.,-/.1/ -2 i g i:;' -4

l ~"o ,s i •• <:i ~ -'2 i ll r"

-4 / ·2 /I./·· .:19>' ..

·2 -3 -2 -1 () 1 2 3 -3 -2 _, 0 1 2 3 qnorm qnorm Figure 42: Q-Q plots of Frequency Asymmetry (M14)

.J ·2 ·1 0 1 2 3 ""' p-,·· /')""·' /,,~, "' i,_ "' ;!i § =; ~ .. i'i ~ li·~\W/>BRI¢>::./ :n}.< :1 :'/llofii'i ,, .../ <·RUM'.·· a § ..,,..,. i .;·~ i 2 '·' ~ m ,,I ?? '5 ! ,,;· I ·I • ,,,,.! ./ I ·2 -3 -2 -1 "-0 1 2 3 .J ·2 -1 0 1 2 3 qnorm qnorm Figure 43: Q-Q plots of Relative Time of Peak Overall Intensity (M18)

.3 -1 ·1 0 1 2 3 J...... ,.J....,-- i l l ,, .·,:'.;sr. ·1r.tP~<(~

'

~ (}<~ l'l u. ,___,,l/./~I 15 .2 /' ~ -1 § ~ l1 § .i·:·::.·:·aRK"''''' •·'l :.•ii .:1R0R· ·.:·r-- :1 :RUMY. ;'l ! ll J

-·" ,~~"'?" I .,__,-· I . __.,,r·

·2 -J ·2 ·1 0 1 2 3 -J -2 -1 G 1 2 3 qnorm qnorm Figure 44: Q-Q plots of Frequency of Peak Overall Intensity (M20)

107 .) •2 •t I} 1 2 J ! ' i______L__L '(, 'TMP

-10 IO

"' I.. _...r t"' & "' ~ ,____ ,,r (!'. .J ~· ,..'!llllt•• '•'!·-'.Ci!ibiL "'JjJJM. ': "' ' ~· ~ to ~

,_

,__.,,,,.~·' I ,,__,,..,~.- ~~~;-:-.~

.2 -3 -2 ·1 0 1 2 ) -J -2 -1 0 1 2 )

qnorm Qn

..J -2 -1 0 1 2 J '-'--L_L_' 0 !'SOti! +· -·TMP"' ·'"-' 2tl 2tl

15

15 10 ;:: :;;; I i ~" (!'. .')!'.T!ii::!ll!!K;'.;!'{• .,•'.lt!T'71::;r-R0R •'· ., ·. "!'!/· . 'Rini!!" · ! 20 i IS ~ ~ ,5

10 oil J. , _,,' ,,,,,,,,,.,... ~------.r ,~ ,--- -2 -3 -2 -I 0 1 2 3 -3 -2 -1 0 1 2 3

qnorm Ql1

..J -2 -1 I) 1 2 3 12 iK".TC0'1;E+ ;·;+: 'S• "'JMIFi'C. •;; •' 12 10 10

~ "' .., ,' 1 51 _ ___; __ ___.,,,,,;:- ti: 6 ~ ,__,/ ~ · -.,,J"iBRK' ·'' ·· i:'!':i'.! '.\ROR·· :C.';i,E".ii\' ('lIDM!f· IS ~ 12 ~ 4 ,. c.:i: ~o g ,!j .. ____/ , ~..---""'"" ,,...-""'' I ·---···

-2 .3 ·2 -1 0 1 2 J .) ·2 -1 (I 1 2 ) qnorm qnorm Figure 47: Q-Q plots of FM Rate (M23)

108 -3 -2 -1 0 1 2 J +-<;/;: -SOK-• ~~P'.-

15 " " ~ .. ~ 10 ~ ~ ~ ·-s>''> .,.. ~ i~~'.,:"tSPk'' "'' -"'-ROO:"'- ;-,.-----~"----- ~.. i 15 ;; . ;

-2 -3 -2 -1 0 I 2 3 -3 -2 -1 0 1 2 3 Cl'OfTTl qoorm Figure 48: Q-Q plots of FM Rate Variation (M24)

..J -2 -1 0 1 2 J

•'' 0 I I "' ~ /' "' g ,~'l/'-/ i ' ~ 2 -r ~~H?Sji;:•:· ----:"RORL: c-• f· --•;-.;,'ROM"•_..,_ - l:;, {l ~.

,_/ ,~~ ,_,~-·

-2 -J -2 -1 0 1 2 3 .3 -2 -1 0 1 2 3 qrorm qoorm Figure 49: Q-Q plots of Overall Entropy (M26)

-l -2 -1 0 1 2 3 '''.U!:SllK''." i!_E,V ·::;·soc.,~::, ,_, .:' ~f'..;_ ~~~ ·>::ThlP·~~: :~<- !~·

..--1 /. ~ 0--- ,,/\~/ i ~ -5 ~ ~ i ! •:Uiit:'-' ,.p: ~ ,,,,.,-----~" ~~,- .. ,• -5

-2 -3 -2 -1 0 1 2 3 .3 -2 -t ti 1 2 3

109 -3 -2 -1 0 1 2 3 I "o',AAK ,A:<;:,~nn ).·. TMP'

•' / ~ 2 ~ ,,, / -2 ~ /I./' { £ ~ ~ >··liRil::·' '<1100' '''<•l -·-,-·< iliJlli;,· i 0 a ll <1l. ll -2 / j,/''" /,•' -2 ..,,,,e

-3 -2 -1 1l 1 2 3 -3 -2 -~ 0 1 2 3 qnorm qnorm Figure 51: Q-Q plots ofUpsweep Fraction (M28)

110 Box PLOTS OF SCALED VARIABLES FOR IDENTIFYING OUTLIERS

0 ll -

0

II -

0 0 0 0 4 - 0 0 0 0 0 €10 0 0 0 0 .- 0 0 0 0 ,--,0 I ,- 0 0,- ,- I I I I 0 . I 2-i I-., I I 'T' I -r--, I 8 0 I -., I I I I I I I I 1-r-r 0 I I I I I I I I I I I I I 0 1 I I I I I I I I I I I I 8 I 0 I H~RBR8B+Q8888-~-Q+$ _i_ ...L. : -'- : I -L : : : -'- _!_ 8 : I I I 0 I I ..J.... 0 I I ~ ..J...... J.... I I -2--t ~ J_ €1 ..J.... 0 0

0 -4 I 0 0

I ,.. ..., ... <:":I 0 ... ~ .... ~ .. e 'i 1i 1ii m. "' ~ ::r e 0 ~ ' "- " .. .s "' ::;:.. "' .. .. !I .... "' ::;:"' .. "' ,g ::;: e 0 "- "- a. :3" ~ I- ..... ~ ::J , 12..... Figure 52: Boxplot of scaled variables for Blow AcoustiC parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

111 0 0

6 - 0

0 0 0 0 4 - 0 0 0 0 0 0 @ GI @ 0 0 -r -. 1::1 I I 0 A t t 6-rT" 0 0 2 - " -rt t Qt t 0 I I I ,.- I I Qg.,-0 t t 8 t t t t 0 0 -1L 1-r-1 I I I I I ,- :T:::: : : : : § :-r "Tit-;

0 -

$$9$8Bf-L t t t B~WO~Q~o~$$-Lt t..L. 0 t I -1..-'-8 ..L. t 0-1... -'- -'- -2 -I 0 8 0 0 0

0 -4 -j 0

-6 - I 0

I I I I I I I I I I I I I I I I I I ... ( :!"' :! :!""' :ii: :ii: :!"' "' N "'N ..... N"" N CT D" c: .c '!: :! :ii: :! :! :ii: :! :ii: "' "'::< :ii: e. !! 0 'S E"' E i:r E t ..... e E E ~ c <> ... .. E ., ti , 1ii"' 'j <>. .,m E ...~ ~ 3 i= <:"' ... §. a: i! a: a: ., E ..._ & ::l "CJ c: 0 i. :! .. Q,. <::: <>. c: ~ '!i'l"' :! '15 0. 3 <>. c '5.. u .!!! if ~ a: Ill"' :::> m.. t er 0 ~ a: ... ~· .s E" "' ~ ... :a .. 0. .. E c... "Cl. :a"' "' 0 .... :::> i= ~ :ii "- il :::> """- Figure 53: Boxplot of scaled variables for Bark Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

112 £II

s~xc-J\ atp uo a.rn s1a+aur1:utid alp Jo san1t?A (paZfit?UIJou) pa1ti:>S ·s~xti-x atp uo a1ti s1a+aurti.rnd ::i9sno::iy .nm-H JOJ sa1qt:pt:A paJt:;Js 10 lOJdxog :ps a.1n~!.i

,.. ..., c: .., ..,, "0 ~ .,,.c: '"Cl ;;; s:: ~ ,l!: .. s:: ~ .<> .. 3 tD c: >E ;o .,, 0 a ·9' .. a. b ::0 "" .. .. "'O s"' IJ;l ai' n ::r 0 "t:r s .,, .. !!!. s: ~ ii. ~ !' ... t:: "'a. ll: ... Q ::0 ;; ;o a ii " 2 ~ ~ ...... " ~ s iil a !!le !a il. .. ~ iil 3 a " i .,, ...... :::t.."' "' .. ., :;; ::; 3' .a 3' <> 3 !ij; c;· ii "'' ~ 0 ... 0 "' .... "' "' ... '"'

8 8 e 0 0 z·

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0

0 0

0 0 6 -

0 0 0 0 0 0 0 4 - 0 0 0 0 0 0 0 0 0 0 0 ~ 0 "'T" 0 I 0 ~ I Q I 0 ~0: :~ 0 Q 2 - I -,- I I I I I "'T""'T" I I I I IR fl "'T" I I 1-r-1 I I I I~ ~ 0 I

!) -

Oa~aga+Q8B~$$8e~+$I -L ....1.... I I I ..J.... -L I ....L. ....L. -L I I I I I 8 I I ....1.... I 0 -'2 - l_ § 0

0

-4 ~

0

0 -!! - I I I I I I I I I I I I I I I I I I

,.,_ 0 "'It -g l5 ~ c ~ .. "' Q. c 0 ~ "' 15 :a: ~ Q. & if .. t "' 0 "' ~ .... l 3 ~ '5 CT ' " :::> ' .3 .. ... ::.:. ., "~ E E .. 2 "" "' J:I. I f=: ;:: ::;; .... I "- ~ ~ ::> ::> x Figure 55: Boxplot of scaled variables for Rumble Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

114 0 0

0 s 1 0 0 0 0 6 ~ 0 0 0 0 0 0 0 0 0 0 0 0 0 4 -I 0 0 e 0 1§1 0 0 0 0 0 0 0 0 0 0 0 0 0 ... 8 8 0 0 -r § 0 0 0 0 i~: o® -9-: 0 •o 2 - 0 : : 1§1 i 0 I ~ I I -Y- g0 -:-I :I :I I J ! 0 :I -.I ...I -r- I ,- I I I I I I I I I () - $8$8~$f $~8~$D~i$$8 I I -L I I I I I I I I I I I : I I I I ~~ ~ ~~~ ~I I -2 - e o v : 0 § ~ 0 8 0 0 0 -4 -I 0 I I I I I I I I I I I I I I I I I I ... It> .... al 0 ... 0 ;:; N ...... "'~ :a :a ""~ ~ ~ "" N "'N N "'N "' r:r C' c: ~ .. 'C ~ ~ ~ ~ :a"' :a ~ :a :a"' :a"' l'! 3 E C" E I- "- ... .. ~ ~ <::: 0 e ...... :!2 .. Ci 15 10 Cl. J! ~ s ;:: c ....e c: c: ~ c: ~., E .."' ... ti e::I .., <::: "0 i c: ~ e .. :a .. .Cl. <: 1i 1i ... s 0 (J t .. t ~ :a ~ Cl. "'- .. '5"' ., 'E 0 ~ c: .... c: ,. II> '5 t:r .:9 ::J .>< Cl. '!l a_ :a ..Cl. "' "" ~" .l ~· e 0 .... ::J I- I- :a"' u. i ::J """CL Figure 56: Boxplot of scaled variables for Squeak Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

115 to 0

0 0 0 0 0 0 0 0 0 0 0 0 ~ 8 0 0 0 0 9 0 0 0 0 0 0 0 8 0 0 8 0 0 ~ 0 o 0 0 ..I e 0 0 0 I • @ I I 0 0 0 0 ~ e 0 ! Jo 8 T .- ! o 8 lo I I •I I 1 0 TI t I ~-I 1-?~ T "T" I -r I I I -r I I I I I I I

0

8$8~~8fl_.L-'- $9898B~1~~$I l...L..1-L I I ...J.... I -LIll I I I 181 I ...J.... I * 0 ...J.... S ~ ~ 0 0 0 8 0 0 0 0

-!i g 0 I I I I

.,, ,... 0 0 ., CD· c ., ..... ::;:"" :e ::;:"' :e :e N N "' "' N "'N :e"' .,. c .::: i i ::;: "' "'::;: "" "' "'::;: :: .. :e :eer :e ..... :e ~ ~ ~ c u I!?"'" l!! :2 E E ...... Q. ... .s <.>"' e . 0 c 0 ~ "' "'Q. "' ~· f ct: ~ s m .,, '6"' Jf,,. I! 0 ~ c:: .... ' 3' .s ::>· "' .. .. ::;: ..a. ::;: _g .. e ~ c.. ..<>. ::> I- ...... s :a I;.. 0 "" ~ ... :::> ""0.. Figure 57: Boxplot of scaled variables for Squeal Acoustic parameters are on the x-axis. Scaled (normalized) values of the parameters are on the y-axis

116 lII

spro-A: aip uo a.m s1aiaurn.md aipJO sanp:?A (paznmurnu) pap:?;)S ·s~xu-x aip uo a1u s1aiaurn.md ;)nsno;)y 1adwn.1i .IOJ sa1q'Bp'BA ps JO io1dxog :ss a.1nlj!.!I ,.."'D c ..., ..., .... "::! c -a "'D 0 ;;J s: .., O> 3 3 s: rA s: ~ .0 .. ~ m c ..., 0 a ..CL .Ei" ::lJ "'" "O ""C i!L ~ "'("j ... ::i 0 .,, ""C " !!i. s: l!i. ~ ~ ::r 0 . Q. ~ "' a .. ~ "' c: -~ "..., s: ::0 ii .,, ::r .." d .. -I ..., iil .. -a "' !!i. l!i. ~ 3 ;;J 3 ~ ~ 0 ..:> "' .,, "'=i 3 .0 3 ".. "j;j; ., ii ii .. .. 3 ..Q ... ::r ::J ..Q s: ..,:s: 1:..., ~ :s: ~ :s: s: :s: s: s: :s: :s:.. s: :s: s: s: s: ..... ~ 0 c _, . "'""' "' '...'"' "' '"' "' "' ... "' "' "' ... "' I I I I I I I I I I I I I I I I I I

0

0 0

8 0

§ 0 ~I ~I """I ~I ~I 8 ~ ~ ~I I 1 T I:: I ,-..,-II":: $$~-~9~~~~~t~~8~S~ - 0 I I () ii- I : I I : l_ :· I I I I l I 0~~ ~ t!l 1 1--1.....L I U I I I I --1... 0 t 0 @ 8 ~ ..L 0 ; 8 8 --1... s --1... g o e o o o 8 @ o o S 0 0 0 0 0 0 0 0 0 0 0

0 0

0 APPENDIX B: CLASSIFICITON TREE RAW DATA OUTPUT

lolHIF~~'7;1

CIClfliJ;l1C.~""

~

vFQC11F...;.w .. 1m $QI( 1''\ln.'161!\W ,__ ["~ ·-+;; r~ ~ 'i;T';:;i.,""!' ... l . .tt'MJGi~\r..r- ~ IUol ®':l!1"'111 t.'\l.-0.:C.1<1 tvnm!artll .£rln~'tl-ll Figure 59: Classification tree final model - plots

Table 12: Classification tree final model - raw data outout [l] 0.1220096 > surnmary(mod) Call: rpart(formula calltype - ., data params) n= 836

CP nsplit rel error xerror xstd 1 0.27638191 0 1.0000000 1.0000000 0.02188309 2 0.22948074 1 0.7236181 0.7236181 0.02420221 3 0.16247906 2 0.4941374 0.4941374 0.02314367 4 0.05276382 3 0.3316583 0.3400335 0.02076694 5 0.03182580 5 0.2261307 0.2428811 0.01833773 6 0.01340034 6 0.1943049 0.2211055 0.01766021 7 0.01005025 7 0.1809045 0.2093802 0.01727080 8 0.01000000 8 0.1708543 0.2043551 0.01709825

Node number 1: 836 observations, complexity param=0.2763819 predicted class=SQL expected loss=0.7141148 class counts: 59 105 57 197 239 179 probabilities: 0.071 0.126 0.068 0.236 0.286 0.214 left son=2 (614 obs) right son=3 (222 obs) Primary splits: LowerFreqM3 < 675.0855 to the left, improve=l63.3921, (0 missing) FMRateVarM24 < 0.0055 to the left, improve=l49.4871, (0 missing) MedianFreqMll < 561.782 to the left, improve=l00.2804, (0 missing) PkOverallFM20 < 530.3175 to the left, improve= 99.7010, (0 missing) DurationM5 < 0.4395 to the right, improve= 92.1828, (0 missing) Surrogate splits: PkOverallFM20 < 1302.758 to the left, agree=O. 868, adj=0.505, (0 split) MedianFreqMll < 1318.912 to the left, agree=0.859, adj=0.468, (0 split) DurationM5 < 2.4235 to the left, agree=0.787, adj=0.198, (0 split) AMRateVarM22 < 0.007 to the right, agree=0.785, adj=0.189, (0 split) TimeConcentM9 < 1.6775 to the left, agree=0.782, adj=0.180, (0 split)

Node number 2: 614 observations, complexity param=0.2294807 predicted ciass=SQK expected loss=0.6840391 class counts: 59 105 57 194 29 170 probabilities: 0.096 0.171 0.093 0.316 0.047 0.277 left son=4 (452 obs) right son=5 (162 obs) Primary splits: FMRateVarM24 < 0.0055 to the right, improve=128.59100, (0 missing) AMRateVarM22 < 0.0215 to the right, improve=l20.66540, (0 missing) PkOverallFM20 < 530.3175 to the left, improve= 97.40044, (0 missing) 118 MedianFreqMll < 561.782 to the left, improve= 97.01046, (0 missing) BandwidthM6 < 2616.284 to the left, improve= 94.27497, (0 missing) Surrogate splits: AMRateVarM22 < 0.0215 to the right, agree=0.879, adj=0.543, (0 split) DurationM5 < 0.287 to the right, agree=0.824, adj=0.333, (0 split) UpswpFracM28 < 34.3755 to the right, agree=0.800, adj=0.241, (0 split) AMRateM21 < 4.0295 to the left, agree=O. 796, adj=0.228, (0 split) TimeConcentM9 < 0.1265 to the right, agree=0.787, adj=0.191, (0 split)

Node number 3: 222 observations predicted class=SQL expected loss=0.05405405 class counts: 0 0 0 3 210 9 probabilities: 0.000 0.000 0.000 0.014 0.946 0.041

Node number 4: 452 observations, complexity param=0.1624791 predicted class=TMP expected loss=0.6238938 class counts: 58 105 57 33 29 170 probabilities: 0.128 0.232 0.126 0.073 0.064 0.376 left son=8 (237 obs) right son=9 (215 obs) Primary splits: UpperFreqM4 < 1866.577 to the left, improve=91.40859, (0 missing) MedianFreqMll < 642.902 to the left, improve=89.47726, (0 missing) BandwidthM6 < 1361.89 to the left, improve=87.85287, (0 missing) PkOverallFM20 < 468.347 to the left, improve=85.20569, (0 missing) FMRateVarM24 < 0. 011 to the right, improve=77.18512, (0 missing) Surrogate splits: BandwidthM6 < 1361.89 to the left, agree=0.985, adj=0.967, (0 split) MedianFreqMll < 680.2285 to the left, agree=0.889, adj=0.767, (0 split) EntropyM26 < 95.3805 to the left, agree=0.863, adj=0.712, (0 split) PkOverallFM20 < 468.347 to the left, agree=0.858, adj=0.702, (0 split) LowerFreqM3 < 241.5455 to the left, agree=O. 827, adj=0.637, (0 split)

Node number 5: 162 observations predicted class=SQK expected loss=0.00617284 class coilllts: 1 0 0 161 0 0 probabilities: 0.006 0.000 0.000 0.994 0.000 0.000

Node number 8: 237 observations, complexity param=0.05276382 predicted class=ROR expected loss=0.5738397 class counts: 46 101 56 19 11 4 probabilities: 0.194 0.426 0.236 0.080 0.046 0.017 left son=l6 (83 obs) right son=l7 (154 obs) Primary splits: · DurationM5 < 0.8675 to the left, improve=33.89742, (0 missing) MedianTimeM7 < 0.37 to the right, improve=28.32799, (0 missing) PkOverallFM20 < 170.0515 to the left, improve=23.31349, (0 missing) LowerFreqM3 < 384.906 to the left, improve=22.73504, (0 missing) AMRateVarM22 < 0.3255 to the left, improve=22.52987, (0 missing) Surrogate splits: MedianTimeM7 < 0.4005 to the left, agree=0.920, adj=0.771, (0 split) AMRateM21 < 1.1595 to the right, agree=O. 869, adj=0.627, (0 split) FMRateM23 < 1.1525 to the right, agree=0.865, adj=0.614, (0 split) TimeConcentM9 < 0.2685 to the left, agree=0.857, adj=0.590, (0 split) UpperFreqM4 < 1184.326 to the right, agree=O. 7 85, adj=0.386, (0 split)

Node number 9: 215 observations, complexity param=0.01005025 predicted class=TMP expected loss=0.227907 class counts: 12 4 1 14 18 166 probabilities: 0.056 0.019 0.005 0.065 0.084 0.772 left son=18 (17 obs) right son=19 (198 obs) Primary splits: UpswpFracM28 < 36.3945 to the left, improve=ll.40962, (0 missing) FMRateVarM24 < 0.011 to the right, improve=ll.36771, (0 missing) UpperFreqM4 < 3016.089 to the left, i~prove=ll.33500, (0 missing) TimeConcentM9 < 1.318 to the right, improve=l0.79894, (0 missing) BandwidthM6 < 2616.284 to the left, improve=l0.19102, (0 missing) Surrogate splits: TimeConcentM9 < 0.0305 to the left, agree=0.930, adj=0.118, (0 split) 119 TimeAsyrnmMlO < -0 .2115 to the left, agree=O .930, adj=O .118, (0 split) DurationM5 < 0.1325 to the left, agree=0.926, adj=0.059, (0 split) FMRateM23 < 15.4305 to the right, agree=0.926, adj=0.059, (0 split)

Node number 16: 83 observations, complexity param=0.0318258 predicted class=BRK expected loss=0.4939759 class counts: 42 11 0 19 9 2 probabilities: 0.506 0.133 0.000 0.229 0.108 0.024 left son=32 (55 obs) right son=33 (28 obs) Primary splits: LowerFreqM3 < 325.017 to the left, improve=21.43904, (0 missing) PkOverallFM20 < 539.676 to the left, improve=l8.54354, (0 missing) AMRateVarM22 < 0.0215 to the right, improve=l8.52093, (0 missing) MedianFreqMll < 669.679 to the left, improve=l8.18897, (0 missing) DurationM5 < 0.3615 to the right, improve=13. 89051, ( O missing) Surrogate splits: MedianFreqMll < 662.2645 to the left, agree=0.940, adj=0.821, (0 split) PkOverallFM20 < 539.676 to the left, agree=0.940, adj=0.821, (0 split) DurationM5 < 0.3615 to the right, agree=0.904, adj=0.714, (0 split) UpperFreqM4 < 1372.742 to the left, agree=0.867, adj=0.607, (0 split) AMRateM21 < 2.7815 to the left, agree=0.867, adj=0.607, (0 split)

Node number 17: 154 observations, complexity param=0.05276382 predicted class=ROR expected loss=0.4155844 class counts: 4 90 56 0 2 2 probabilities: 0.026 0.584 0.364 0.000 0.013 0.013 left son=34 (92 obs) right son=35 (62 obs) Primary splits: PkOverallFM20 < 170.0515 to the right, improve=28.37610, (0 missing) AMRateVarM22 < 0.3255 to the left, improve=24.03463, (0 missing) FMRateVarM24 < 0.0695 to the left, improve=22.60592, (0 missing) LowerFreqM3 < 22.5425 to the right, improve=l8.17635, (0 missing) EntropyM26 < 25.117 to the right, improve=l5.06470, (0 missing) Surrogate splits: MedianFreqMll < 181.817 to the right, agree=0.857, adj=0.645, (0 split) LowerFreqM3 < 35.6645 to the right, agree=O. 786, adj=0.468, (0 split) AMRateVarM22 < 0.087 to the left, agree=0.753, adj=0.387, (0 split) FMRateVarM24 < 0.0195 to the left, agree=O. 740, adj=0.355,. (0 split) UpperFreqM4 < 425.1125 to the right, agree=0.714, adj=0.290, (0 split)

Node number 18: 1.7 observations predicted class=SQK expected loss=0.4705882 class counts: 0 0 0 9 5 3 probabilities: 0.000 0.000 0.000 0.529 0.294 0.176

Node number 19: 198 observations predicted class=TMP expected loss=0.1767677 class counts: 12 4 1 5 13 163 probabilities: 0.061 0.020 0.005 0.025 0.066 0.823

Node number 32: 55 observations predicted class=BRK expected loss=0.2363636 class counts: 42 11 0 0 1 1 probabilities: 0.764 0.200 0.000 0.000 0.018 0.018

Node number 33: 28 observations, complexity param=0.01340034 predicted class=SQK expected loss=0.3214286 class counts: 0 0 0 19 8 1 probabilities: 0.000 0.000 0.000 0.679 0.286 0.036 left son=66 (19 obs) right son=67 (9 obs) Primary splits: FMRateVarM24 < 0. 011 to the right, improve=ll.007940, (0 missing) AMRateVarM22 < 0.0215 to the left, improve= 9.385714, (0 missing) EntropyM26 < 97.3315 to the right, improve= 8.058442, (0 missing) BandwidthM6 < 1012.061 to the right, improve= 4.919048, (0 missing) PkOverallRel TM18 < 56.2775 to the left, improve= 4.500000, (0 missing) Surrogate splits: AMRateVarM22 < 0.0215 to the left, agree=0.964, adj=0.889, (0 split) 120 EntropyM26 < 80.201 to the right, agree=0.929, adj=0.778, (0 split) PkOverallRelTM18 < 56.2775 to the left, agree=0.857, adj=0.556, (0 split) BandwidthM6 < 973.779 to the right, agree=0.821, adj=0.444, ( 0 split) MedianFreqMll < 693.84 to the right, agree=0.821, adj=0.444, (0 split)

Node number 34: 92 observations predicted class=ROR expected loss=0.1630435 class counts: 0 77 11 0 2 2 probabilities: 0.000 0.837 0.120 0.000 0.022 0.022

Node number 35: 62 observations predicted class=RUM expected loss=0.2741935 class counts: 4 13 45 0 0 0 probabilities: 0.065 0.210 0.726 0.000 0.000 0.000

Node number 66: 19 observations predicted class=SQK expected loss=O class counts: 0 0 0 19 0 0 probabilities: 0.000 0.000 0.000 1.000 0.000 0.000

Node number 67: 9 observations predicted class=SQL expected loss=0.1111111 class counts: 0 0 0 0 8 1 probabilities: 0.000 0.000 0.000 0.000 0.889 0.111

·--rt.•

--··»...... 10. !,j~

•..*-.w -~·------·----~~:::o""""'i .... L _[J_

Figure 60: Classification tree validation model - plots

Table 13: Classification tree validation model - raw data output (1] 0.2368421 > summary(new) Call: rpart(formula calltype - ., data= parru~s) n= 836

CP nsplit rel error xerror xstd 1 0.2763819 0 1.0000000 1.0000000 0.02188309 2 0.2294807 1 0.7236181 0.7252931 0.02420020 3 0.1624791 2 0.4941374 0.4974874 0.02317902 4 0.1000000 3 0.3316583 0.3400335 0.02076694

Node number 1: 836 observations, complexity param=0.2763819 predicted class=SQL expected loss=0.7141148 class counts: 59 105 57 197 239 179 probabilities: 0.071 0.126 0.068 0.236 0.286 0.214 left son=2 (614 obs) right son=3 (222 obs) Primary splits: LowerFreqM3 < 675.0855 to the left, improve=163.3921, (0 missing) FMRateVarM24 < 0.0055 to the left, improve=149.4871, (0 missing) MedianFreqMll < 561.782 to the left, improve=l00.2804, (0 missing)

121 PkOverallFM20 < 530.3175 to the left, improve= 99.7010, (0 missing) DurationM5 < 0.4395 to the right, improve= 92.1828, (0 missing) Surrogate splits: PkOverallFM20 < 1302.758 to the left, agree=0.868, adj=0.505, (0 split) MedianFreqMll < 1318.912 to the left, agree=0.859, adj=0.468, (0 split) DurationM5 < 2.4235 to the left, agree=0.787, adj=0.198, (0 split) AMRateVarM22 < 0.007 to the right, agree=0.785, adj=0.189, (0 split) TimeConcentM9 < 1.6775 to the left, agree=0.782, adj=0.180, (0 split)

Node number 2: 614 observations, complexity param=0.2294807 predicted class=SQK expected loss=0.6840391 class counts: 59 105 57 194 29 170 probabilities: 0.096 0.171 0.093 0.316 0.047 0.277 left son=4 (452 obs) right son=5 (162 obs) Primary splits: FMRateVarM24 < 0.0055 to the right, improve=128.59100, (0 missing) AMRateVarM22 < 0.0215 to the right, improve=120.66540, (0 missing) PkOverallFM20 < 530.3175 to the left, improve= 97.40044, (0 missing) MedianFreqMll < 561.782 to the left, improve= 97.01046, (0 missing) BandwidthM6 < 2616.284 to the left, improve= 94.27497, (0 missing) Surrogate splits: AMRateVarM22 < 0.0215 to the right, agree=0.879, adj=0.543, (0 split) DurationM5 < 0.287 to the right, agree=0.824, adj=0.333, (0 split) UpswpFracM28 < 34.3755 to the right, agree=0.800, adj=0.241, (0 split) AMRateM21 < 4.0295 to the left, agree=0.796, adj=0.228, (0 split) TimeConcentM9 < 0.1265 to the right, agree=0.787, adj=0.191, (0 split)

Node number 3: 222 observations predicted class=SQL expected loss=0.05405405 class counts: 0 0 0 3 210 9 probabilities: 0.000 0.000 0.000 0.014 0.946 0.041

Node number 4: 452 observations, complexity param=0.1624791 predicted class=TMP expected loss=0.6238938 class counts: 58 105 57 33 29 170 probabilities: 0.128 0.232 0.126 0.073 0.064 0.376 left son=8 (237 obs) right son=9 (215 obs) Primary splits: UpperFreqM4 < 1866.577 to the left, improve=91.40859, (0 missing) MedianFreqMll < 642.902 to the left, improve=89.47726, (0 missing) BandwidthM6 < 1361. 89 to the left, improve=87.85287, (0 missing) PkOverallFM20 < 468.347 to the left, improve=85.20569, (0 missing) FMRateVarM24 < 0.011 to the right, improve=77.18512, (0 missing) Surrogate splits: BandwidthM6 < 1361. 89 to the left, .agree=O. 985, adj=0.967, (0 split) MedianFreqMll < 680.2285 to the left, agree=0.889, adj=0.767, (0 split) EntropyM26 < 95.3805 to the left, agree=0.863, adj=0.712, (0 split) PkOverallFM20 < 468.347 to the left, agree=0.858, adj=0.702, (0 split) LowerFreqM3 < 241.5455 to the left, agree=0.827, adj=0.637, (0 split)

Node number 5: 162 observations predicted class=SQK expected loss=0.00617284 class counts: 1 0 0 161 0 0 probabilities: 0.006 0.000 0.000 0.994 0.000 0.000

Node number 8: 237 observations predicted class=ROR expected loss=0.5738397 class counts: 46 101 56 19 11 4 probabilities: 0.194 0.426 0.236 0.080 0.046 0.017

Node number 9: 215 observations predi~ted class=TMP expected loss=0.227907 class counts: 12, 4 1 14 18 166 probabilities: 0.056 0.019 0.005 0.065 0.084 0.772

122 APPENDIX C: PRINCIPAL COMPONENT ANALYSIS DATA - RAW DATA OUTPUT

Table 14: PrinciEal ComEonent Anall'.sis - raw data outEut Standard deviation 2.1032890 1.8106105 1.31698516 1.15100919 1.11924231 Proportion of Variance 0.2460624 0.1823465 0.09647373 0.07368938 0.06967798 Cumulative Proportion 0.2460624 0.4284088 0.52488258 0.59857195 0.66824993

Loadings: Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 LowerFreqM3 -0.29 -0.14 -0.35 -0.17 0.14 UpperFreqM4 -0.42 0.25 0.15 DurationM5 -0.52 BandwidthM6 -0. 38 0. 35 0. 20 MedianTimeM7 -0.51 0.18 TimeConcentM9 -0.51 -0.11 0.12 TimeAsyrnmM10 -0.32 0. 35 MedianFreqMll -0.44 0.15 FreqAsyrnmM14 0.33 0.20 -0.23 PkOverallRelTM18 0.28 -0.47 PkOverallFM20 -0.38 -0.10 -0.24 -0.13 0.14 AMRateM21 0.24 -0.34 0.17 AMRateVarM22 0.23 0.28 0.12 0.42 FMRateM23 0.24 -0.25 0.10 FMRateVarM24 0.21 0.24 0.46 EntropyM26 -0. 38 0.24 0.15 UpswpMeanM27 0.17 -0.54 -0.29 UpswpFracM28 -0.16 0. 31 -0.48 -0.25

broken.stick(18) j E (j) [1, l 1 0.194172671 [2' l 2 0.138617115 [3' l 3 0.110839338 [4, l 4 0.092320819 [5, l 5 0.078431930 [6, l 6 0.067320819 [7, l 7 0.058061560 [8, l 8 0.050125052 [9'} 9 0.043180608 [10,] 10 0.037007768 [11, l 11 0.031452212 [12,] 12 0.026401707 [13,] 13 0.021772078 [14,] 14 0.017498574 [15, l 15 0.013530320 '[16,] 16 0.009826616 [17,] 17 0.006354394 [18,J 18 0.003086420 Comp.l Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Proportion of Variance 0.2460624 0.1823465 0.09647373 0.07368938 0.06967798 0.05750879

123 APPENDIX D: ANALYSIS OF SIMILARITY (ANOSIM) DAT A OUTPUT

Table 15: Pair-wise ANOSIM plots for 15 pair-wise tests Pairs are noted above each -plot BRKv.ROR BRKv.RUM

R:: 0??7 P:: 7Q P-

~ ' ' ~ : : ; § 0 § s g 0 or iii ~

0 0 18~ Between SRK ROR RUM SQK SOL TMP Between BRK ROR ~RUM SQK SOL TMP

BRKv.SQK BRKv.SQL

0 R=0~1<1 P=<•"'nn1 R::::fllAll P::

0 § ~ B 0 + 9 0 i Between BRK' ROR RUM SQK SOL TMP Between BR~ ROR RUM SQK SQL TMP BRKv.TMP RORv.RUM

D: ""'' P: <111101 ~= n.H7 P=

g ~ ~ g ' ~

0 § ~ 6 g ~ ~

8 ~ 0 ~ ~ 0 i j Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SQL TMP

124 RORv.SQK RORv.SQL

~ § ~ l ~ § ~ l'l 9 ~ 0 8 ~ 0 8 0 r ~ s Between BRK ROR RUM SOK SOL TMP Bef'.Neen BRK ROR RUM SQK SQL TMP

RORv.TMP RUMv.SQK

g ~ ~ 1 Ii

0 0 i t ~ g r ~ ·@ T-T ~ R § ' 0 8 0 !t ~ Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SOL TMP

RUMv.SQL RUMv.TMP

O: n;:qs:.. P= c:"nn• g ~ ~ 8 ~ ~ ~ ~ ~ ~ 0 r ~ 0 t ! 9 Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SOL TMP

125 SQKv. SQL SQKv. TMP

g - ____n~ 0 i ::: i g ;;\ :il i g f g "' y 18 y H i ~ ~ l 0 .'! 8 8 -.- Between BRK ROR RUM SQK SQL TMP Between BRK ROR RUM SQK SOL 1MP

SQLv.TMP

!J 8 ~ g~ BetWeen BRK ROR RUM SQK SQL TMP

126 APPENDIX E: POWER SPECTRA OF RUMBLES AND NOISE

Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth 105,--~~..--~~~~~--,~~~,.-~~-,.-~~----. 105~~~~~~~~~~~~~1

100 100 95

~ 00 95 ~ ,! 851- ~ i;g" 00 v 90 75 85 70

80 65 200 400 600 800 1000 1200 100 200 3lO 400 v500 600 700 800 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

95.-~~..--~~~~~--,~~~,.-~~-,.-~~----. 100~~..--~,.-~~~~~~~~~~~~..--~~~

90

80

~ g~ ~ i ig"

70

65

ffi 200 400 600 800 ~ = 30 2000 4000 GOOD 8000 10000 12000 14000 16000 18000 frequency, Hz frequency, Hz Figure 61: Power spectra of Rumble and background noise at the Oregon Zoo (example 1) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

127 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth 100 100

n 95~ I~ 95

90 ~

~ ~ . . 90 ~ a5 ~ i " 80 ! rg" !g- 85 75 ~ 00 70

65 75'-~'--~--'-~~-'-~~-'-~~'--~--'-~~-'-__J 100 200 300 400 500 600 700 800 900 1000 1100 100 200 300 400 500 600 700 frequency, Hz frequency. Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

90 90~~~~~~~~~~~~~~~~~~~~~-

00

~ ~ ~ 80 .0 c ~ ~ ~ 60 i § !g- 75 rg" so

70 40

65'--~~'--~~-'-~~--'-~~~'--~~-'-~~--' 3)'-~-'-~-'-~-'-~-'-~--'-~~'-~'--~-'-~-'-~~ JJO 400 600 BOO 1000 1JJO JJOO 4000 6000 8000 10000 12000 14000 16000 18000 frequency, Hz frequency, Hz. Figure 62: Power spectra of Rumble and background noise at the Oregon Zoo (example 2) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

128 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

~ ~ ~ 5 Ell !fi 70 ~~ 75 65 70 60 ~

55 65~~~~~~~~~~~~~~~~~-~~~~ 100 200 300 400 500 600 700 800 OCIJ 1000 1100 5IJ M ~ D B m B a s 500 frequency, Hz frequency. Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

100 100~~~~~~~~~~~~~~~~~~~~

95 90

90 Ell 85 70 80

75

70

65

60

55i 20 :m 400 500 800 1000 1200 2000 4000 6000 8000 10000 12000 14000 16000 1BOOJ frequency, Hz frequency, Hz Figure 63: Power spectra of Rumble and background noise at the Oregon Zoo (example 3) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

129 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

65

~ 65f ""u ~ 60 i 60f § 55 aj 55' ~\ "" 50

SOf 45

40 45 ~

35 ~ 40 100 200 300 400 500 600 700 800 900. 1000 1100 50 100 150 200 250 300 350 400 frequency, Hz frequency.Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

70~~~~~~~~~~~~~~~~~~~~ 10~~~~~--~~~~~~~~~~~~~~

65 65 60

55 -g 55 g ""g 50 j 50 i 45 § rg" 45 !g- 40

40

35 25

3J 200 400 600 800 1000 1200 20 2000 4000 6DOO 8000 10000 12000 14000 16000 18000 frequency, Hz frequency, Hz Figure 64: Power spectra of Rumble and background noise at the Elephant Nature Park (example 1) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

130 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

oo~~~~~~~~~~~~~~~~~~~~~ so~~~~~~~~~~~~~~~~~~~~~

75

70

65

~ ~ 60 65 ~ i 55 § 60 !fl. 50

45 55 40 50 35

45 30 200 400 --- 40 60 BO 100 120 140 160 180 200 220 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

oo~~~~~~~~~~~~~~~~~~~---, oo~~~~~~~~~~~~~~~~~~~---,

75 70 70

65 60

50

45

40

35 ~: 30'---'-~-'-~-'----''-----'~-'-~-'-~-'--~'-----'~--'--~~ 10 100 200 300 400 500 600 700 800 900 1000 1100 ~ ~ = ~1~1=1~1~1~ frequency, Hz frequency. Hz Figure 65: Power spectra of Rumble and background noise at the Elephant Nature Park (example 2) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

131 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

BO so~~~~~~~~~~~~~~~~~~~~~~

75

70

65 ... g 60 ] 65 ~ 60 j 55 ~ ~ ~· 51) ~· 55 45 50 40

35 45

30 40 100 200 300 400 500 600 700 800 900 100J 1100 50 100 150 200 250 300 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

70.----~~~~~~~~~~~~~~.----~~~~ 70.----~~~~~~~~~~~~~~~~~~~

65 60

50 ~ e so i 45 iii iii 40

35

30

25 10~~~~~~~~~~~~~~~~~~~~~~ 100 200 300 400 500 600 700 800 900 1000 1100 2000 4000 6000 8000 10000 12000 14000 1600) 18000 frequency, Hz frequency, Hz Figure 66: Power spectra of Rumble and background noise at the Elephant Nature Park (example 3) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

132 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth Ill Ill

75 75 70

65 70

~ ~w ~ 60 g 65 ~ ~ i 55 i § 60 ; ro !ll.

45 55 40 50 35

~ 45 100 200 300 400 500 600 700 800 900 1000 11 DO 50 100 150 200 250 300 350 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

Ill.-~~.-~~~~~~~~~~~~~~~--, BO.-~.-~.-~.-~.-~.-~.-~.-~.-~.---,

75 70 70 60

~~~--'~~~~~~-'-~~~~~---'-~~--' 10 200 400 600 800 1000 1200 =~ ~~1=1=1~1~1~ frequency, Hz frequency, Hz Figure 67: Power spectra of Rumble and background noise at the Elephant Nature Park (example 4) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

133 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

.,, 65 .,,. i ~ 65 i SJ i " 60 ~.55 ill-

55 50

45 50

40 45L-LC.---'-.---'-.--~.---'-.---'-.---L.----'.--.--L-.---'-J 100 200 300 400 500 600 700 800 9()0 1000 1100 50 100 150 200 250 300 350 400 450 500 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

80.--.--.--.--.--.--~.--.--~.--.--.--~.--.--~.--.-----. 80.--.--~.--~.--~.--~.--~.--~.--~.--~.--~~

70

60 .,, 'g 65 . ~ I!! ~ 50 -s 60 § i 40 ~- 55 ill-

30 50

45 20

40 10 200 400 600 800 1000 1200 2000 4000 6000 8000 10000 12000 14000 16000 18000 frequency, Hz frequency, Hz Figure 68: Power spectra of Rumble and background noise at the Royal Elephant Kraal (example 1) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

134 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

85.---~~~~~~~~~~~~~~~~~~~~

65

60

55

50

\f\~ 45

35 40 - 200 400 600 Ellll 1000 1200 10U 200 300 400 500 600 frequency, Hz frequency, Hz

Nmse.. . 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

70 m~~~~~~~~~~~~~~~~~~~~~

60

50 ~ ~ ~ 55 u e ~ i 50 40 5 i !g. 45 ~ 30 40 20 35

30 100 200 DJ 400 500 600 700 Ellll 900 1000 1100 w ~ ~ =~1~1=uooo1=1~ frequency, Hz 1requem:y, Hz Figure 69: Power spectra of Rumble and background noise at the Royal Elephant Kraal (example 2) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

135 Rumble: 14 Hz to 1.2 Hz Rumble: call bandwidth

85~~~~~~~~~~~~~~-....~~~~-.... as~-....~~~~~~~~~~~~~~~~~-....

00

.,, g 65 !! ~ ro 60 § !g. 55 55

50 50

45 45 40 40 ~I./

35 35 100 200 300 400 500 600 700 800 900 1000 1100 100 200 300 400 500 600 700 800 900 frequency, Hz frequency, Hz

Noise: 14 Hz to 1.2 kHz Noise: 14 Hz to 20 kHz

65~·~~~~~~~~~~~~~~~~~~

60

55

50 i: 50 ~ 45 'i 45 § ~ 40

35

30

20 25'-----x~~~o;!;n-----'--_J200 400 600 800 11XXJ 1200 21XXJ 4000 6000 8000 10000 12000 14000 16000 18000 frequency, Hz frequency, Hz Figure 70: Power spectra of Rumble and background noise at the Royal Elephant Kraal (example 3) Frequency is on the x-axis. Power (dB) is on the y-axis. Power is referenced to the same value.

136