INTERSPEECH 2012 ISCA Archive ISCA's 13th Annual Conference http://www.isca-speech.org/archive Portland, OR, USA September 9-13, 2012 Lenition of /d/ in spontaneous Spanish and Catalan Miquel Simonet1, José I. Hualde2, Marianna Nadeu2 1Dept. of Spanish & Portuguese, University of Arizona, Tucson, Arizona, USA 2Dept. of Spanish, Italian & Portuguese, University of Illinois at Urbana-Champaign, Illinois, USA [email protected], [email protected], [email protected] first approach is to classify spectrographic tokens into categories Abstract (and subcategories) on the basis of visual inspection [5]. The The present study explores the acoustics of /d/ in two corpora of problem with this approach is that it forces researchers to divide Spanish and Catalan spontaneous speech. Three acoustic metrics a continuum into categories whose boundaries can only be fuzzy. were developed as indexes of articulatory weakening. The A second approach is to use RMS (Root Mean Square) energy findings suggest that variations in the implementation of /d/ measurements as indicative of degree of constriction. The energy result from gradient modulations in constriction degree on a output or intensity of the segment can be considered an acoustic unimodal statistical-acoustic distribution. The preceding segment correlate of degree of constriction: open, vowel-like realizations is a strong predictor of the weakening of Catalan and Spanish /d/. of /b d g/ are expected to have much greater acoustic intensity Index Terms: Spanish, Catalan, lenition, phonetics, phonology. than very closed stop realizations. Acoustic intensity, however, can only be measured in 1. Introduction relative terms, since it greatly depends on linguistically irrelevant factors such as generally softer or louder elocution, distance to The systematic lenition or “spirantization” of voiced plosives is a the microphone, etc. Several normalization methods have been phonological areal feature of most of the Iberian Peninsula. In developed that have used different window sizes as references to the languages of this area, voiced plosives are systematically express acoustic intensity in relative terms. For instance, realized as voiced approximants, called “fricatives” in earlier references that have been used include the whole word in which work, in some contexts. In the Iberian Peninsula, this the consonant appears [11], the VCV sequence (i.e., the average phenomenon has been reported for Spanish [e.g., 1, 2], Catalan of the intensity maxima of the two surrounding vowels) [12] and [e.g., 3, 4], Galician [5], Basque [6], and as an optional the peak intensity of the following vowel [e.g., 10]. phenomenon in Northern-Central Portuguese [7]. An entirely different approach has been to explore velocity Spirantization in these languages has been described as the curves derived from acoustic intensity contours [13]. In this complementary distribution between two categorically distinct procedure, minima (highest falling velocity, i.e. from previous allophones, stop and spirant. For instance, standard Spanish /b d segment to target consonant) and maxima (highest rising g/ are said to have two types of allophones in complementary velocity, i.e. from target consonant to following segment) values distribution: stop allophones in utterance-initial position and are extracted from velocity curves. The actual minima and after a homorganic nasal (or lateral in the case of /d/, but not /b/ 10.21437/Interspeech.2012-50 maxima of the intensity curves are ignored; the velocity values or /g/) and spirant allophones elsewhere [1, 2]. However, it has are key. This procedure is based on the premise that more been pointed out that there is substantial variation in the constricted consonants will have more extreme velocity values constriction of these sounds in different word tokens [8, 9, 10]. due to more abrupt segment-to-segment transitions [13]. In This suggests that, even if realizations can be classified as laboratory speech a good correlation has been found between presenting or lacking complete occlusion, the phenomenon may these acoustic measurements and articulation [14], although not be better understood as a continuum of constriction degrees. in contexts such as /nl/ and /ld/, where the acoustic output may In line with this observation, it has been reported that, in not reflect the presence of complete occlusion [15]. addition to the preceding context (i.e., the standard phonological The present study examines the effect of the preceding description), speaking rate, lexical stress, morpheme identity and segment on the degree of constriction of /d/ in two comparable the specific timbre of the surrounding vowels condition degree of corpora of conversational speech, one for Spanish and one for constriction in at least some of these sounds, [e.g., 10]. Yet, Catalan, using three different acoustic measurements. Majorcan while some degree of gradience has indeed been attested, this Catalan shows less extensive lenition of /b/ than Iberian Spanish could be an artifact of a continuous distribution that might turn [16]. For /d/, the difference appears to be smaller, but in the same out to be bimodal if examined in detail. A bimodal gradient direction [17, p. 322]. Our findings suggest that Iberian (Catalan distribution could call for the application of a phonological rule and Spanish) spirantization is a non-categorical weakening acting at some higher level of abstraction; that is, two process conditioned (in part) by the degree of constriction of the categorically distinct allophones. Is spirantization unimodally preceding segment. This gradualness extends to the intervocalic gradient in Iberian Romance? context. Even though this is the prototypical context for A question that arises is how to obtain information on the spirantization, we find that higher preceding vowels condition articulatory parameter of degree of consonantal constriction from more constricted allophones than lower vowels. the non-linear, indirect information that sound waves provide. A INTERSPEECH 2012 1416 2. Method statistical treatment. For the statistical analysis we applied mixed effects regression modeling, using individual speaker and lexical 2.1. Recordings item as random factors. The comparisons that we make are slightly different for the two languages due to differences in the A total of 40 speakers were recruited for participation in an vowel inventories. interactive task with one of the authors. The data were collected on the island of Majorca, Spain, a Catalan-Spanish bilingual 3. Results speech community. Twenty (10 females) Catalan-dominant bilinguals were recorded in Catalan, and 20 (10 females) 3.1. Spanish Spanish-dominant bilinguals were recorded in Spanish. Language dominance was assessed by means of a language The dataset comprised a total of 241 Spanish /d/ tokens from 20 background questionnaire and a series of accent rating tasks. speakers. The data were visually explored in order to see whether The participants were asked to engage in an interactive uni- or bimodal distributions were found. Kernel density plots speech game. Speech was recorded through a head-mounted were obtained for all three acoustic metrics, ID, ST and MV. microphone into a solid-state recorder (44.1 kHz, 16-bit). For Density plots display the estimated probability density function each participant, we have approximately 15 minutes of speech. of a continuous random variable. These plots are similar to histograms but provide a continuous estimate of the distribution 2.2. Acoustic analyses of a variable, rather than grouping observations into discrete bins. The kernel density plots clearly showed unimodal Three different acoustic metrics were used as correlates of distributions for all three metrics (see Figure 1). degree of constriction. One measurement, intensity difference (ID), is the difference in acoustic intensity between the Spanish - Maximum Rising Velocity Spanish - Spectral Tilt consonant minimum and the following vowel’s maximum: the 0.06 more open the constriction, the smaller the difference with 0.0012 respect to the following vowel. A pass Hann band filter (500- 0.05 10,000 Hz) was applied to the signal prior to obtaining the 0.04 0.0008 intensity curve. This filtering procedure excludes the effects of 0.03 f0. This is hypothesized to maximize the difference in this metric Density Density 0.02 between approximants and voiced stops. Therefore, if a slight 0.0004 difference in constriction degree between the two sounds exists, 0.01 this method will increase the likelihood of detecting it. 0.00 0.0000 A second metric, spectral tilt (ST), is obtained by calculating -500 0 500 1000 1500 -30 -20 -10 0 10 N = 241 Bandwidth = 95.69 N = 241 Bandwidth = 2.021 the difference in acoustic energy in the portion corresponding to Spanish - Intensity Difference /d/ (hand-segmented) between a low frequency band (50-500 Hz) 0.04 and a high-energy band (500-5000 Hz): the more open the constriction, the greater amount of energy is expected in the 0.03 higher zone of the spectrogram, resulting in a smaller difference in energy between the lower and the higher frequency bands. 0.02 In a third metric we calculate the first difference of the Density intensity curve in steps of 1 ms. Then we extract the maximum 0.01 value (MV) in the velocity curve between the intensity minimum corresponding to /d/ and the maximum corresponding to the 0.00 following vowel. This measurement reduces possible effects of -10 0 10 20 30 40 50 variation in the intensity of the following vowel and focuses on the abruptness of the transition between consonant and vowel: Figure 1. Density plots for three acoustic indexes of constriction more weakened consonants have a less abrupt transition degree in 241 tokens of Spanish /d/ recorded by 20 speakers. irrespective of the time from the intensity minimum to the intensity maximum and, therefore, a lower MV than more For the statistical explorations we classified the tokens as a constricted consonants. This procedure is a simplified adaptation function of the preceding segment in the following groups: after of the method described above [13].