DOCSLIB.ORG

What Do Phone Embeddings Learn About Phonology?

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
What Do Phone Embeddings Learn About Phonology? What do phone embeddings learn about Phonology? Sudheer Kolachina Lilla Magyar [email protected] [email protected] Abstract resentations using human judgement datasets (Ba- roni et al., 2014; Levy et al., 2015). Asr and Jones Recent work has looked at evaluation of phone (2017) use artificial language experiments to study embeddings using sound analogies and corre- the difference between similarity and relatedness lations between distinctive feature space and embedding space. It has not been clear what in evaluating distributed semantic models. Phone aspects of natural language phonology are embeddings induced from phonetic corpora have learnt by neural network inspired distributed been used in tasks such as word inflection (Sil- representational models such as word2vec. fverberg et al., 2018) and sound sequence align- To study the kinds of phonological relation- ment (Sofroniev and C¸ oltekin¨ , 2018). Silfverberg ships learnt by phone embeddings, we present et al.(2018) show that dense vector representa- artificial phonology experiments that show tions of phones learnt using various techniques are that phone embeddings learn paradigmatic re- able to solve analogies such as p is to b as t is to lationships such as phonemic and allophonic distribution quite well. They are also able X, where X = d. They also show that there is a to capture co-occurrence restrictions among significant correlation between distinctive feature vowels such as those observed in languages space and the phone embedding space. with vowel harmony. However, they are un- Our goal in this paper is to understand better the able to learn co-occurrence restrictions among evaluation of phone embeddings. We argue that the class of consonants. significant correlation between distinctive feature 1 Introduction space and phone embedding space cannot be auto- matically interpreted as the model’s ability to cap- Over the last few years, distributed represen- ture facts about the phonology of natural language. tation models based on neural networks such Since many distinctive features tend to be pho- as word2vec (Mikolov et al., 2013a) and netically based, natural classes denoted by these GloVe (Pennington et al., 2014) have been of features capture phonetic facts as well as phono- much importance in speech and natural language logical facts. For example, the feature [±long] processing (NLP). The word2vec technique is denotes the distinction between long and short a shallow neural network that takes a text corpus vowels, which is a language-independent phonetic as input and outputs a vector space containing all fact. But, whether this distinction is a phonolog- unique words in the text. The dense vector rep- ical fact varies from language to language. It is resentations of words induced using word2vec important to make this distinction between pho- have been shown to capture multiple degrees netic facts and phonological facts when evaluating of similarities between words. Mikolov et al. phone embeddings for their learning of phonology. (2013a,b) show that word embeddings can solve In this paper, we propose an alternative method- word analogy questions and sentence completion ology to evaluate word2vec’s ability to learn tasks. Mikolov et al.(2013b) show that word phonological facts. We define artificial languages embeddings represent words in continuous space, with different kinds of phoneme-allophone dis- making it possible to perform algebraic opera- tinctions and co-occurrence restrictions and study tions, such as vector(King) − vector(Man) + vec- how well phone embeddings capture these rela- tor(Woman) = vector(Queen). Considerable atten- tionships. Several interesting insights regarding tion has been paid to evaluating these vector rep- the relationship between phonetics and phonol- 160 Proceedings of the 16th Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 160–169 Florence, Italy. August 2, 2019 c 2019 Association for Computational Linguistics ogy, the role of distinctive features and the task of dition to distinctive features in phonology, there distinctive feature/phoneme induction accrue from are also phonetic features that describe the artic- our experiments. ulatory and acoustic properties of phones (Lade- foged and Johnson, 2010). However, in practice, 2 Background and Related work there is considerable overlap between phonologi- cal distinctive features and phonetic features. This One major difference between words and phones already poses an interesting question about the is that while words are meaningful units in lan- nature of the relationship between phonetics and guage, phones have no meaning in themselves. phonology, which as we will see, is relevant to the However, as with words, there are clear patterns of evaluation of phone embeddings. organization of individual phones in a language. Next, let us examine the notion of correlation One well-known pattern in phonology is the dis- between distinctive feature space and phone em- tinction between contrastive and complementary bedding space to evaluate phone embeddings as distribution. Two phones are said to be in con- proposed by Silfverberg et al.(2018). Pair-wise trastive distribution if they occur in the same con- featural similarity is estimated using a metric such text and create a meaning contrast. For example, as Hamming distance or Jaccard index applied to b and k occur in word-initial position and create feature representations of phones. Pair-wise con- a contrast in meaning, such as in bæt versus kæt. textual similarity is estimated as cosine similar- This is why they are considered distinct phonemes ity between phone embeddings induced using a h in the language. On the other hand, p and p never technique like word2vec. The correlation be- occur in the same context, which is referred to as tween pairwise featural similarity and pairwise being in complementary distribution. Since they contextual similarity is estimated using Pearson’s are phonetically related, they are considered allo- r or Spearman’s ρ. The value of this correla- phones, variants of the same underlying phoneme. tion is shown for a number of languages in ta- The notions of contrastive and complementary dis- ble1. Data for Shona and Wargamay are taken tribution are purely based on context. They can from Hayes and Wilson(2008) 1. Similar datasets be considered instances of paradigmatic similar- were constructed for Telugu and the Vedic va- ity discussed in the distributed semantic literature. riety of Sanskrit2. For English, the CMU pho- Allophony also involves the notion of phonetic netic dictionary was used with a feature represen- similarity. Another pattern in natural language tation based on Parrish(2017) with some minor phonology is that of co-occurrence restrictions. extensions. The word2vec implementation in A well-known example is homorganic consonant the Gensim toolkit (Rehˇ u˚rekˇ and Sojka, 2010) was clusters. For example, in nasal plus stop clusters, used to induce phone embeddings using the fol- the nasal must have identical place of articulation lowing parameters- CBOW, dimensionality of 30, to the following stop. Yet another example of window size of 4, negative sampling of 3, mini- co-occurrence restriction in phonology is the phe- mum count of 5, learning rate of 0:05. We use nomenon of vowel harmony. In some languages, CBOW which predicts the most likely phone given a word can only have vowels which agree with re- a context of 4 phones in either direction as this is spect to certain features, such as backness, round- intuitively similar to the task of a phonologist. It ing or height. Co-occurrence restrictions can be would be interesting to compare CBOW and Skip- considered to be instances of syntagmatic similar- gram architectures and also, study the effect of dif- ity whereby words that frequently occur together ferent parameters on this correlation between dis- form a syntagm (phrase). Again, most types of co- tinctive feature space and phone embedding space. occurrence restrictions involve phonetic similarity. However, this is not the goal of our study. In this The traditional method to describe phones paper, we restrict our attention to the linguistic sig- in phonology is in terms of distinctive fea- nificance of this correlation. tures (Jakobson et al., 1951). Distinctive features All languages in Table1 show a significant pos- allow phones to be grouped into natural classes, itive correlation between distinctive feature space which are established on the basis of participa- 1 tion in common phonological processes. They https://linguistics.ucla.edu/people/ hayes/Phonotactics/index.htm#simulations allow for generalizations about phonotactic con- 2Datasets and code available at https://github. texts to be captured in an economical way. In ad- com/skolachi/sigmorphoncode 161 Language Size Pearson Spearman meaning, when embeddings of two phones show English 135091 0:589 0:612 Shona 4395 0:431 0:575 high similarity, it is not clear if it is an instance of Telugu 19627 0:349 0:350 paradigmatic similarity (phonemic relationship) or Wargamay 5910 0:411 0:428 syntagmatic similarity (co-occurrence restriction). Vedic 45334 0:351 0:285 English 4000 0:129 0:161 Feature Class Shona 4000 0:507 0:533 -high a0,a1,a2,aa1 Telugu 4000 0:202 0:206 +high i0,i1,i2,ii1,u0,u1,u2,uu1,w,y +long aa1,ii1,uu1 Wargamay 4000 0:219 0:387 -long a0,a1,a2,i0,i1,i2,u0,u1,u2 Vedic 4000 0:146 0:159 +back a0,a1,a2,aa1,u0,u1,u2,uu1,w -back i0,i1,i2,ii1,y -approximant N,b,d,g,j,m,n,nj Table 1: Correlation between distinctive feature space +approximant R,a0,a1,a2,aa1,i0,i1,i2,ii1,l,r,u0,u1,u2,uu1,w,y and embedding space, all values significant (p < 0:01) -sonorant b,d,g,j +sonorant N,R,a0,a1,a2,aa1,i0,i1,i2,ii1,l,m,n,nj,r,u0,u1,u2,uu1,w,y +syllabic a0,a1,a2,aa1,i0,i1,i2,ii1,u0,u1,u2,uu1 -syllabic N,R,b,d,g,j,l,m,n,nj,r,w,y +main a1,aa1,i1,ii1,u1,uu1 and embedding space.
Load more

Recommended publications
  • Writing & Language Development Center Phonetic Alphabet for English Language Learners Pin, Play, Top, Pretty, Poppy, Possibl
    Writing & Language Development Center Phonetic Alphabet f o r English Language Learners A—The Phonetic Alphabet (IPA) is a system of phonetic symbols developed by linguists to represent P each of the wide variety of sounds (phones or phonemes) used in spoken human language. This includes both vowel and consonant sounds. The IPA is used to signal the pronunciation of words. Each symbol is treated separately, with examples (like those used in the dictionary) so you can pronounce the word in American English. Single consonant sounds Symbol Sound Example p p in “pen” pin, play, top, pretty, poppy, possible, pepper, pour t t in “taxi” tell, time, toy, tempted, tent, tender, bent, taste, to ɾ or ţ t in “bottle” butter, writer, rider, pretty, matter, city, pity ʔ or t¬ t in “button” cotton, curtain, kitten, Clinton, continent, forgotten k c in “corn” c in “car”, k in “kill”, q in “queen”, copy, kin, quilt s s in “sandal” c in “cell” or s in “sell”, city, sinful, receive, fussy, so f f in “fan” f in “face” or ph in “phone” gh laugh, fit, photo, graph m m in “mouse” miss, camera, home, woman, dam, mb in “bomb” b b in “boot” bother, boss, baby, maybe, club, verb, born, snobby d d in “duck” dude, duck, daytime, bald, blade, dinner, sudden, do g g in “goat” go, guts, giggle, girlfriend, gift, guy, goat, globe, go z z in “zebra” zap, zipper, zoom, zealous, jazz, zucchini, zero v v in “van” very, vaccine, valid, veteran, achieve, civil, vivid n n in “nurse” never, nose, nice, sudden, tent, knife, knight, nickel l l in “lake” liquid, laugh, linger,
    [Show full text]
  • FACTORS AFFECTING PROFICIENCY AMONG GUJARATI HERITAGE LANGUAGE LEARNERS on THREE CONTINENTS a Dissertation Submitted to the Facu
    FACTORS AFFECTING PROFICIENCY AMONG GUJARATI HERITAGE LANGUAGE LEARNERS ON THREE CONTINENTS A Dissertation submitted to the Faculty of the Graduate School of Arts and Sciences of Georgetown University in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Linguistics By Sheena Shah, M.S. Washington, DC May 14, 2013 Copyright 2013 by Sheena Shah All Rights Reserved ii FACTORS AFFECTING PROFICIENCY AMONG GUJARATI HERITAGE LANGUAGE LEARNERS ON THREE CONTINENTS Sheena Shah, M.S. Thesis Advisors: Alison Mackey, Ph.D. Natalie Schilling, Ph.D. ABSTRACT This dissertation examines the causes behind the differences in proficiency in the North Indian language Gujarati among heritage learners of Gujarati in three diaspora locations. In particular, I focus on whether there is a relationship between heritage language ability and ethnic and cultural identity. Previous studies have reported divergent findings. Some have found a positive relationship (e.g., Cho, 2000; Kang & Kim, 2011; Phinney, Romero, Nava, & Huang, 2001; Soto, 2002), whereas others found no correlation (e.g., C. L. Brown, 2009; Jo, 2001; Smolicz, 1992), or identified only a partial relationship (e.g., Mah, 2005). Only a few studies have addressed this question by studying one community in different transnational locations (see, for example, Canagarajah, 2008, 2012a, 2012b). The current study addresses this matter by examining data from members of the same ethnic group in similar educational settings in three multi-ethnic and multilingual cities. The results of this study are based on a survey consisting of questionnaires, semi-structured interviews, and proficiency tests with 135 participants. Participants are Gujarati heritage language learners from the U.K., Singapore, and South Africa, who are either current students or recent graduates of a Gujarati School.
    [Show full text]
  • Phones and Phonemes
    NLPA-Phon1 (4/10/07) © P. Coxhead, 2006 Page 1 Natural Language Processing & Applications Phones and Phonemes 1 Phonemes If we are to understand how speech might be generated or recognized by a computer, we need to study some of the underlying linguistic theory. The aim here is to UNDERSTAND the theory rather than memorize it. I’ve tried to reduce and simplify as much as possible without serious inaccuracy. Speech consists of sequences of sounds. The use of an instrument (such as a speech spectro- graph) shows that most of normal speech consists of continuous sounds, both within words and across word boundaries. Speakers of a language can easily dissect its continuous sounds into words. With more difficulty, they can split words into component sounds, or ‘segments’. However, it is not always clear where to stop splitting. In the word strip, for example, should the sound represented by the letters str be treated as a unit, be split into the two sounds represented by st and r, or be split into the three sounds represented by s, t and r? One approach to isolating component sounds is to look for ‘distinctive unit sounds’ or phonemes.1 For example, three phonemes can be distinguished in the word cat, corresponding to the letters c, a and t (but of course English spelling is notoriously non- phonemic so correspondence of phonemes and letters should not be expected). How do we know that these three are ‘distinctive unit sounds’ or phonemes of the English language? NOT from the sounds themselves. A speech spectrograph will not show a neat division of the sound of the word cat into three parts.
    [Show full text]
  • English Teachers' Mastery of the English Aspiration And
    Arina Isti’anah - English Teachers’ Mastery of the English Aspiration and Stress Rules ENGLISH TEACHERS’ MASTERY OF THE ENGLISH ASPIRATION AND STRESS RULES Arina Isti’anah Sanata Dharma University [email protected] Abstract This paper tries to observe the English teachers’ awareness and representation of the English aspiration and stress rules. The research purposes to find out whether or not the teachers are aware of the English aspiration and stress rules, and to find out how the teachers represent the English aspiration and stress rules. Based on the analysis, it can be concluded that the teachers’ awareness of the English aspiration and stress rules is very low. It is indicated with the percentage which equals to 44% and 48% for English aspiration and stress rules. In representing the English aspiration and stress rules, the teachers face the problems in producing aspiration in the pronunciation, placing the right stress and pronouncing three and four X in the coda position. There are two reasons affecting the teachers’ awareness of the English aspiration and stress rules namely exposure and L1 influence. Artikel ini bertujuan untuk meneliti kesadaran dan representasi aturan aspirasi dan tekanan oleh guru bahasa Inggris. Penelitian ini bertujuan untuk menjelaskan apakah guru bahasa Inggris mempunyai kesadaran atas aturan aspirasi dan tekanan dalam bahasa Inggris, dan untuk menunjukkan bagaimana guru Bahasa Inggris mewujudkan aturan aspirasi dan tekanan dalam pelafalan mereka. Berdasarkan analisis yang dilakukan, dapat disimpulkan bahwa kesadaran guru Bahasa Inggris atas aturan aspirasi dan tekanan dalam Bahasa Inggris masih sangat rendah. Hal tersebut ditunjukkan oleh rendahnya prosentase dalam perwujudan aturan aspirasi dan tekanan: 44% dan 48%.
    [Show full text]
  • A Deep Generative Model of Vowel Formant Typology
    A Deep Generative Model of Vowel Formant Typology Ryan Cotterell and Jason Eisner Department of Computer Science Johns Hopkins University, Baltimore MD, 21218 ryan.cotterell,eisner @jhu.edu { } Abstract and its goal should be the construction of a univer- sal prior over potential languages. A probabilistic What makes some types of languages more approach does not rule out linguistic systems com- probable than others? For instance, we know pletely (as long as one’s theoretical formalism can that almost all spoken languages contain the vowel phoneme /i/; why should that be? The describe them at all), but it can position phenomena field of linguistic typology seeks to answer on a scale from very common to very improbable. these questions and, thereby, divine the mech- Probabilistic modeling also provides a discipline anisms that underlie human language. In our for drawing conclusions from sparse data. While work, we tackle the problem of vowel system we know of over 7000 human languages, we have typology, i.e., we propose a generative proba- some sort of linguistic analysis for only 2300 of bility model of which vowels a language con- them (Comrie et al., 2013), and the dataset used in tains. In contrast to previous work, we work di- rectly with the acoustic information—the first this paper (Becker-Kristal, 2010) provides simple two formant values—rather than modeling dis- vowel data for fewer than 250 languages. crete sets of phonemic symbols (IPA). We de- Formants are the resonant frequencies of the hu- velop a novel generative probability model and man vocal tract during the production of speech report results based on a corpus of 233 lan- sounds.
    [Show full text]
  • Phonics TRB Coding Chart
    Coding Charts The following coding charts briefly explain vowel and spelling rules, syllable-division patterns, letter clusters, and coding marks used in Saxon’s phonics programs. Basic Coding TO CODE USE EXAMPLE Accented syllables Accent marks noÆ C ’s that make a /k/ sound, as in “cat” K-backs |cat C ’s that make a /s/ sound, as in “cell” Cedillas çell Combinations; diphthongs Arcs ar™ Digraphs; trigraphs; quadrigraphs Underlines SH___ Final, stable syllables Brackets [fle Long vowel sounds Macrons nO Schwa vowel sounds (rhymes with vowel sound in “sun,” as Schwas o÷ (or ) in “some,” “about,” and “won”) Short vowel sounds Breves log Sight words Circles ≤are≥ Silent letters Slash marks mak´ Affixes Boxes work ingfl Syllables Syllable division lines cac\tus Voiced sounds Voice lines hiß Vowel Rules RULE CODING EXAMPLE A vowel followed by a consonant is short; code it logcatsit with a breve. An open, accented vowel is long; code it with a nOÆ mEÆ íÆ gOÆ macron. AÆ\|cor™n OÆ\p»n EÆ\v»n A vowel followed by a consonant and a silent e is long; code the vowel with a macron and cross out the nAm´ hOp´ lIk´ silent e. An open, unaccented vowel can make a schwa b«\nanÆ\« E\rAs´Æ hO\telÆ sound. The letters e, o, and u can also make a long sound. The letter i can also make a short sound. JU\lŒÆ di\vId´Æ Copyright by Saxon Publishers, Inc. Spelling Rules† RULE EXAMPLE Floss Rule: When a one-syllable root word has a short vowel sound followed by the sound /f/, /l/, or /s/, it is puff doll pass usually spelled ff, ll, or ss.
    [Show full text]
  • Grapheme-To-Phone Using Finite-State Transducers D
    GRAPHEME-TO-PHONE USING FINITE-STATE TRANSDUCERS D. Caseiro, I. Trancoso, L. Oliveira C. Viana INESC-ID/IST CLUL RuaAlves Redol 9, 1000-029 Lisbon, Portugal Av.Prof. Gama Pinto 2, Lisbon, Portugal ABSTRACT Some of the most common approaches to grapheme-to-phone convertion can be compiled to WFSTs,among which are CARTs Severalapproaches have been adopted over the years for [6], and most rule systems, such as two-level [7] and rewriting grapheme-to-phone conversion for European Portuguese: hand- rules [8]. derived rules, neural networks, classification and regression trees, In this work, we first show how we compiled the rules of the etc. This paper describes different approaches implemented as DIXI system to WFSTs (Section 2), we then present data-driven Weighted Finite State Transducers (WFSTs), motivated by their approaches to the problem (Section 3), and finally we combine the flexibility in integrating multiples sources of information and other knowledge-based with the data-driven approaches (Section 4). interesting properties such as inversion. We describe and compare In order to assess the performance of the different methods, we rule-based, data-driven and hybrid approaches. Best results were used a pronunciation lexicon built on the PF (“Portuguˆes Funda- obtained with the rule-based approach, but one should take into mental”) corpus. The lexicon contains around 26000 forms. 25% account the fact that the data-driven one was trained with automat- of the corpus was randomly selected for evaluation. The remaining ically transcribed material. portion of the corpus was used for training or debugging. The sizeofthetraining material for the data-driven approaches was increased with a subset of the BD-P´ublico [9] text corpus.
    [Show full text]
  • Globalphone: Pronunciation Dictionaries in 20 Languages
    GlobalPhone: Pronunciation Dictionaries in 20 Languages Tanja Schultz and Tim Schlippe Cognitive Systems Lab, Karlsruhe Institute of Technology (KIT), Germany [email protected] Abstract This paper describes the advances in the multilingual text and speech database GLOBALPHONE a multilingual database of high-quality read speech with corresponding transcriptions and pronunciation dictionaries in 20 languages. GLOBALPHONE was designed to be uniform across languages with respect to the amount of data, speech quality, the collection scenario, the transcription and phone set conventions. With more than 400 hours of transcribed audio data from more than 2000 native speakers GLOBALPHONE supplies an excellent basis for research in the areas of multilingual speech recognition, rapid deployment of speech processing systems to yet unsupported languages, language identification tasks, speaker recognition in multiple languages, multilingual speech synthesis, as well as monolingual speech recognition in a large variety of languages. Very recently the GLOBALPHONE pronunciation dictionaries have been made available for research and commercial purposes by the European Language Resources Association (ELRA). Keywords: Speech, Text, and Dictionary Resources for Multilingual Speech Processing 1. Introduction More than ten years ago we released a multilingual text With more than 7100 languages in the world (Lewis et al., and speech corpus GLOBALPHONE to address the lack of 2013) and the need to support multiple input and output databases which are consistent across languages (Schultz, languages, it is one of the most pressing challenge for the 2002). By that time the database consisted of 15 languages speech and language community to develop and deploy but since then has been extended significantly to cover more speech processing systems in yet unsupported languages languages, more speakers, more text resources, and more rapidly and at reasonable costs (Schultz, 2004; Schultz and word types along with their pronunciations.
    [Show full text]
  • Collation Sounds Boo
    In a nutshell……… Synthetic Phonics Teaching: The best way to teach the technical skills of reading (decoding) and spelling (encoding) in the English language is to teach the core code knowledge of The English Alphabetic Code in systematic steps and the three core skills of: 1. READING - sound out and blend (synthesise) the sounds (phonemes) represented by the letters and letter groups (graphemes) all-through-the-printed-word, from left to right (e.g. see ‘tray’, say “/t/ /r/ /ai/”, hear and say “tray”). 2. SPELLING - segment (or split up) the smallest identifiable sounds (phonemes) all-through-the-spoken-word (e.g. hear “tray”, identify /t/ /r/ /ai/) and then pull letter/s from memory to spell the word ‘tray’. 3. WRITING - record the correct shapes of the letters or letter groups (graphemes), from left to right, which represent the phonemes identified from segmenting the spoken word from beginning to end. The English Alphabetic Code: We can identify around 44 phonemes (the smallest identifiable sounds in words) in the English language but there are only 26 letters in The Alphabet to represent the 44+ sounds. Single letters and letters combined into letter groups act as code for the sounds, for example; the grapheme ‘ie’ is pronounced /igh/ as in the word ‘tie’. The English Alphabetic Code is complicated by the fact that it has many ‘spelling alternatives’ and ‘pronunciation alternatives’, for example; the grapheme ‘ie’ can also be pronounced /ee/ as in the word ‘chief’. The Alphabetic Code, therefore, needs to be taught explicitly and systematically for both reading and spelling.
    [Show full text]
  • The Phonetics and Phonology of Retroflexes Published By
    The Phonetics and Phonology of Retroflexes Published by LOT phone: +31 30 253 6006 Trans 10 fax: +31 30 253 6000 3512 JK Utrecht e-mail: [email protected] The Netherlands http://wwwlot.let.uu.nl/ Cover illustration by Silke Hamann ISBN 90-76864-39-X NUR 632 Copyright © 2003 Silke Hamann. All rights reserved. The Phonetics and Phonology of Retroflexes Fonetiek en fonologie van retroflexen (met een samenvatting in het Nederlands) Proefschrift ter verkrijging van de graad van doctor aan de Universiteit Utrecht op gezag van de Rector Magnificus, Prof. Dr. W.H. Gispen, ingevolge het besluit van het College voor Promoties in het openbaar te verdedigen op vrijdag 6 juni 2003 des middags te 4.15 uur door Silke Renate Hamann geboren op 25 februari 1971 te Lampertheim, Duitsland Promotoren: Prof. dr. T. A. Hall (Leipzig University) Prof. dr. Wim Zonneveld (Utrecht University) Contents 1 Introduction 1 1.1 Markedness of retroflexes 3 1.2 Phonetic cues and phonological features 6 1.3 Outline of the dissertation 8 Part I: Phonetics of Retroflexes 2 Articulatory variation and common properties of retroflexes 11 2.1 Phonetic terminology 12 2.2 Parameters of articulatory variation 14 2.2.1 Speaker dependency 15 2.2.2 Vowel context 16 2.2.3 Speech rate 17 2.2.4 Manner dependency 19 2.2.4.1 Plosives 19 2.2.4.2 Nasals 20 2.2.4.3 Fricatives 21 2.2.4.4 Affricates 23 2.2.4.5 Laterals 24 2.2.4.6 Rhotics 25 2.2.4.7 Retroflex vowels 26 2.2.5 Language family 27 2.2.6 Iventory size 28 2.3 Common articulatory properties of retroflexion 32 2.3.1 Apicality 33 2.3.2 Posteriority
    [Show full text]
  • Phone Merger Specification for Multilingual ASR: the Motorola Polyphone Network
    Phone Merger Specification for Multilingual ASR: The Motorola Polyphone Network Lynette Melnar and Jim Talley Motorola Human Interface Labs, Voice Dialog Systems Lab, USA E-mail: [email protected], [email protected] ABSTRACT 2. MOTPOLY: OVERVIEW This paper describes the Motorola Polyphone Network MotPoly is a knowledge-based, hierarchically arranged (MotPoly), a hierarchical, universal phone phone merger specification network for shared correspondence network that defines allowable phone multilingual and multi-dialect acoustic modeling. The mergers for shared acoustic modeling in multilingual and organization of MotPoly is language-independent and multi-dialect automatic speech recognition (ML-ASR). phone merger is defined by an internally ranked system of MotPoly’s organization is defined by phonetic similarity relative phonetic similarity, phonological contrastiveness, and other language-independent phonological factors. and phone frequency. MotPoly functions in ML-ASR as a Unlike other approaches to shared acoustic modeling, phone merger framework that constrains data-driven MotPoly can be effectively used in systems where acoustic modeling strategies. Because MotPoly is not computational resources are limited, such as portable biased toward any particular language, language family, devices. Furthermore, it is less constrained by language or language type, it is a universal, static definition of data availability than other approaches. With MotPoly as phone merger and can be used unmodified to specify part of an overall strategy, Motorola’s Voice Dialog likely mergers for an unlimited number of phones from Systems Lab’s ML-ASR team was able to define a set of any collection of languages or dialects. multilingual acoustic models whose size was only 23% of the largest monolingual model set but whose overall MotPoly is of particular value in applications that are performance was higher than the monolingual models by restricted in computational and language resources.
    [Show full text]
  • Phonemes and Allophones of English Consonants
    Rough definition of phoneme • Phoneme (Concise Dictionary of English consonants: Linguistics, Oxford U. Press 1997) Phonemes and Allophones • “The smallest distinct sound unit in a given language: e.g. /»tIp/ in English realizes the Effects related to aspiration and three successive phonemes, represented in ‘devoiced’ voiced sounds and a few spelling by the letters t, i, and p. other issues Phonemic differences vs. Phonemes allophonic differences • Strict, detailed definitions of the term phoneme are • Differences in speech sound that can signal complex differences between two different words are – Not part of this course phonemic differences – Take phonology courses to fight over the details • Other differences in speech sound that are • Rough and ready idea is indispensable for practical phonetics clearly audible are only allophonic – Must make a distinction between phonemic and differences allophonic differences – ‘pronunciation variants’ that cannot signal different words. Representing allophonic Answer: ‘pie, spy, buy’ differences • ‘Broad’ (= coarse-grained) transcription enough Phonemes in ‘/’ (slash or solidus, pl solidi) for phonemic representation marks – Choose simple symbol for a ‘representative’ (allo)phone /p/ /b/ • ‘Narrow’ (= fine-grained) transcription often requires diacritics • Diacritics for stops [p] [pH] [b] [b8] pH - aspirated p p| - ‘p with inaudible release’ (‘unreleased p’) b8 - ‘(partially) devoiced b’ Phones in square brackets Examples ‘Stop.’, ‘Stop!’, Examples: ‘pie, spy, buy’ ‘Stop!!’, ‘Stob!’ • ‘pie’ [»pHaj]
    [Show full text]