Kristine M. Yu

The role of creaky voice in Cantonese tonal perceptiona)

There are few studies on the role of phonation cues in the perception of lexical tones in tonal languages where pitch is the primary dimension of contrast. This study shows that listeners are sensitive to creaky phonation in native tonal perception in Cantonese, a language in which the low falling tone, Tone 4, has anecdotally been reported to be sometimes creaky. First, in a multi-speaker corpus of lab speech, it is documented that creak occurs systematically more often on Tone 4 than other tones. Second, for stimuli drawn from this corpus, listeners identified Tone 4 with 20% higher accuracy when it was realized with creak than when it was not. Third, in a two-alternative forced choice task of identifying stimuli as Tone 4 or Tone 6 (the low level tone) isolating creak from any concomitant pitch cues, listeners had a higher proportion of Tone 4 responses for creaky stimuli. Finally, listeners had more Tone 4 responses for creaky stimuli with longer durations of nonmodal phonation. These results underscore that differences in voice quality contribute to human perception of tone alongside f0. Automatic tonal recognition and clinical applications for tone would benefit from attention to voice quality beyond f0 and pitch.

The role of creaky voice quality in Cantonese tonal perception.

Vance (1976) found a response bias against Tone 4 (mid‐low falling) in a tonal perception experiment in Cantonese where synthesized stimuli varied only in F0, and Vance (1977) explained this by suggesting that creaky voice quality is a redundant cue for Tone 4. Indeed, there is evidence that creaky voice quality plays a role in tonal perception: in Mandarin, a language where creak is well‐known to be a redundant cue for one of the tones, Tone 3 (low fall‐rise), Belotel‐Grenie and Grenie (1994) found that creaky instances of Tone 3 were recognized faster than non‐creaky instances. The effect of creaky voice quality on the perception of Tones 4 and 6 (mid‐low level) in Cantonese will be investigated using a 2AFC identification task of natural stimuli that were elicited in isolation and in connected speech. Variability in creak in the realization of Tone 4 occurred naturally in the elicited stimuli. If creaky voice quality plays a role in tonal perception, we hypothesize that overtly audible creak, as well a...

Contextual tonal variation in level tone languages.

While contextual tonal variation in languages with contour tones has been well studied [Han and Kim (1974); Gandour (1992); Xu (1994)], contextual tonal variation has been less well‐studied in level tone languages [Meyers (1976)]. The acoustic variation of tones due to different tonal contexts in Hausa, Bole (with two tones H and L) and Yoruba (with three tones H, M, and L) will be investigated. The description of cross‐linguistic contextual tonal variation is essential for understanding the task that infants are undertaking in acquiring tonal categories in the face of variance in tonal realization. The focus of the investigation will be how the acoustic variation compares to that in languages with contour tones and if there are general patterns across languages or language‐specific differences in variation.

Investigating tonal spaces using an extension of VoiceSauce voice analysis software

We extended VOICESAUCE (Shue, Keating, and Vicenik, 2009), a Matlab application which provides automated voice measurements over time from audio recordings, to include utilities for command line processing and testing. The command line utilities allow users to access core VOICESAUCE functionality, including batch processing of wave (*.wav) files and parameter manipulation, independently of a graphical user interface. The testing framework provides an automated process for tracking and measuring the effects of manipulating parameter settings across runs. In addition, we modified VOICESAUCE to be compatible with Octave and ported it to Python in order to facilitate use and development from a wider community. We use this software to compare the inclusion of phonation measures in the set of voice source parameters against f0 alone across White Hmong and Cantonese. Phonation, specifically breathy voice, plays a perceptual role in tone identification in both languages; however, in White Hmong, breathy voice is ...

Acoustic representations for tonal phonological categories.

The learnability of tonal categories was investigated by studying how category separability is affected by different phonetic representations of tone. Gauthier et al. [(2007)] found that f 0 velocity contours densely sampled over time were sufficient for near‐perfect categorization of Mandarin tones. Since infant learners begin as “citizens of the world” before developing language‐specific representations of sound categories in response to ambient language input [Kuhl (2004)], cross‐linguistic data were studied. Based on tonal production data from Bole, Igbo, Mandarin, Cantonese, and White Hmong, we found that (1) densely sampled f 0 velocity contours are insufficient for learning tones, but (2) coarse temporal sampling of phonetic features can produce well‐separated tonal categories, and moreover, that (3) phonetic features for tonal representation necessarily extend beyond f 0 to voice quality features, and (4) features for tonal representation are language‐specific. Results (2) and (3) were also suppor...

Speech rhythm and pitch patterns in Bengali: Implications for prosodic acquisition.

Rhythmic timing, measured as durational variability of consonantal and vocalic intervals, is thought to be a cue for phonological chunking in infant acquisition [Ramus et al. (1999)], i.e., for learning stress‐timing (English) versus syllable‐timing (Spanish). Another source of rhythm that has been less studied in infant speech development is “macrorhythm” in pitch contours, regularity in f0 marking of prosodic structure [Jun (2005)]: Head‐marking languages with variable contours on each word (English) are distinguished from edge‐marking languages with a recurring pattern on each word (Korean). Bengali is a language with a moderately constrained syllable structure, and both head‐ and edge‐marking: weak stress and a recurring rise on each content word [Khan, (in press)]. To explore how Bengali can be best categorized in terms of rhythmic timing and macrorhythm, and how speakers may adjust these properties to aid infant language acquisition, we analyzed various durational and intonational measurements in re...

The learnability of language-specific fundamental frequency-based rhythmic patterns

In addition to durational patterns, melodic patterns from variations in f0 contribute to percepts of speech rhythm, too (June 2005, Niebuhr 2009). The overall regularity of these melodic patterns is language-specific. In lexical tonal languages like Mandarin, f0 patterns may be quite variable, as many or all syllables may be associated with a tone; in stress-accent languages like English, f0 variation may be less variable as it is driven by stress and prosodic boundaries (Eady, 1982). In accentual phrase languages like Bengali and French, f0 variation is typically quite regular as it is driven almost entirely by prosodic boundaries in a very constrained way. With only acoustic parameters from the input available, are melodic patterns from these different kinds of languages learnable? And how is learnability affected if speech is directed to young infants? To study this, we recorded parents reading language samples in laboratory speech and in simulated infant directed speech in the languages mentioned abov...

An acoustic and electroglottographic study of Cantonese tone.

Cantonese is a tone language with six tones (traditionally described as high level, high rising, mid level, mid‐low falling, mid‐low rising, and mid‐low level), and also three “stopped” tones in CVC words, not studied here. The interest of this study was (1) how different acoustic cues can be used to classify the different tones, particularly in connected speech and (2) the interaction of voice quality and tone in Cantonese. Twelve speakers were recorded producing minimal pairs as in Wong [(2006)] in isolation, in isolated disyllables, and in sentence‐medial disyllables, where the disyllables ranged over all possible 36 bitones in the language; all speakers also made electroglottographic (EGG) recordings. Acoustic measures were F0, F0′, cepstral peak prominence, and harmonic amplitudes H1* and H2*, H1*‐H2*, H1*‐A1*, H1*‐A2*, H1*‐A3*, and H2*‐H4*. EGG measures were closed quotient and peak of increasing contact. Measures were made automatically using VOICESAUCE and EGGWORKS. Preliminary results indicate va...

Using voice quality to learn non‐native tonal categories.

An artificial language learning experiment will be used to study if voice quality can be used to learn tonal systems. Cross‐linguistically, tone, and voice quality can co‐vary in different ways. For instance, Mazatec (Jalapa de Diaz), has three phonation types (modal, breathy, creaky) fully crossed with three level tone levels [Ladefoged et al. (1988)], while Mandarin has creaky phonation in Tone 3 and Tone 4, which are also distinguished from one another and Tone 1 and 2 by f0 contour [Davison (1991); Belotel‐Grenie and Grenie (2004)]. The goal of the study is to investigate if English listeners unfamiliar with tone languages can use phonation‐type contrasts to learn tonal contrasts in artificial tone languages differing in how tone and voice quality co‐vary. There will be a comparison of how learners generalize from training to novel stimuli for artificial languages where voice quality and tone contrasts are correlated and uncorrelated cues.

The prosody of second position clitic placement and focus in Croatian

Since Browne 1974, the placement of second position clitics in Bosnian/Croatian/Serbian has inspired debate about interaction at the syntax‐phonology interface. The placement of these clitics can alternate quite freely: either after the first phonological word or after the first syntactic constituent. While its generally agreed that prosodic phonology, in addition to morphosyntax, plays a role in clitic placement in Bosnian/Croatian/Serbian, the prosodic patterns for the different placements have not been studied acoustically. In addition, it has been suggested that clitic placement and pragmatic focus may interact (Boskovic 2001), but this has not been systematically studied. We recorded adult Zagreb Croatian speakers producing subject noun phrases with initially stressed trisyllabic adjectives and nouns in transitive sentences. We varied clitic placement (after the first word or first constituent) and focus domain (broad focus, and narrow focus on adjective, noun, both the adjective and the noun, or th...