Understanding Variation of VOT in spontaneous speech Yao Yao UC Berkeley yaoyao@berkeley.edu
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Background Keywords VOT (Voice Onset Time) VOT Variation Spontaneous speech VOT (Voice Onset Time) The duration of time between consonant release and the beginning of voicing of the next vowel Sensitive to speaker and speaking environment close release vowel onset A talk on VOT variation for speaker identification in this conf.
Background What conditions length of VOT? Place of articulation (POA) VOT increases as POA moves backward, i.e. [p]<[t]<[k] Following vowel Speaking rate Age, gender Dialectal background Speech disorders Lung volume Hormone level …
Background Why using spontaneous speech data? Previous results are mostly based on experimental data or read speech. The existence of large-scale transcribed speech corpora makes it possible to study patterns with “naturalistic” data. (Cf. Bell et al. 1999, Gahl in press, Raymond et al. 2006, etc)
Background Experimental data Spontaneous data Controlled content Easy to investigate individual factors Hard to see the general pattern of variation Not necessarily natural speech Spontaneous data Uncontrolled content Need to statistically control for irrelevant factors Provides a general picture of variation More naturalistic. Include factors such as disfluency
Background Purpose of this study Main statistical tool To investigate some of the factors that have been shown to affect VOT in experiments, as well as those that have been proposed to influence spontaneous speech production Main statistical tool Linear regression Adding variables step by step
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Data Buckeye corpus (Pitt et al. 2005) 40 speakers All residents at Columbus, Ohio Balanced in age and gender 1-hr interview Transcribed at word and phone level 19 speakers’ transcriptions were available at the time of this study At the time of this study, 19 speakers’ data were transcribed completely.
Data 2 speakers’ data are used for this study Target tokens F07: Older, female, low speaking rate (4.022 syllables/sec) M08: Younger, male, high speaking rate (6.434 syllabes/sec) Target tokens word-initial transcribed voiceless stops (i.e., [p], [t], [k])
Data Finding point of burst An automatic algorithm is used first. (cf. Yao 2007) >70% of the tokens are checked manually. Error <3.5 ms. Some tokens are rejected by the algorithm for not having significant burst point. 3.03% of F07’s tokens are rejected. 15.85% of M08’s tokens are rejected. Number of tokens F07 M08 Target tokens 231 618 Target tokens with burst point found 210 466
VOT by speaker F07: Mean = 57.41ms, SD = 26.00ms M08: Mean = 34.86ms, SD = 19.82ms
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Preliminary analysis: POA VOT by POA in F07 VOT by POA in M08 Canonical rule: VOT increases when POA moves backward, from lips to the palate. This trend is shown in M08’s data (F(2,463, 14.061) =1.77e-6), but not in F07 (F(2, 207, 3.9925) = 0.01989). p t k p t k
Preliminary analysis: Word class Split the data set into three subsets Content words Function words Other. (e.g. proper names) Number of words of different classes Content Function Other F07 155 47 8 M08 346 104 16
Preliminary analysis: Word class VOT by word class in F07 VOT by word class in M08 Previous literature shows that words of different classes are processed differently. In particular, function words are processed differently than content words. (XXX; YYY) Content words and function words differ greatly in usage frequency. In general, function words are much more often used than content words. Therefore, there could also be a confounding frequency effect. In both speakers, function words are on average shorter than content words, but the variation is vast. F07: F(2, 206, 7.5593) =0.000679 M08: F(2,457, 14.693) =6.541e-07 function content other function content other
Preliminary analysis: word class Word class distinction or general effect of frequency? Obviously word frequency and word class are two related measures, since function words are in general much more frequently used than content words.
Preliminary analysis: word frequency Two frequency measures: Log of Celex frequency Log of Buckeye frequency (speaker-specific) The two measures are highly correlated (r=0.826) Effect: more frequent words have shorter VOT The effect of word class suggests that there might be a more general effect of word frequency: word forms that are used more often tend to be shorter. R^2 indicates how much variance is explained in the regression model. Frequency effect Celex frequency Buckeye frequency p R^2 (%) F07 <0.001 5.1 4.8 M08 4.9 5.9
Word class vs. frequency After factoring out the effect of word class, frequency is no longer significant in F07’s data (p=0.277), but still in M08’s data (p=0.003) This suggests that the above frequency effect in F07 is mainly due to the effect of word class. In other words, we need to factor out the effect of word class if we really want to study the effect of frequency. Previous literature also suggests that content words and function words are processed differently, therefore it’s hard to see homogeneous effect in the overall dataset, that the two must be separated.
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Linear regression model We decide to only model the variation in the content word set F07: 155 tokens M08: 346 tokens Factors investigated POA Word frequency Phonetic context Speech rate Utterance position
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Regression: POA The canonical rule of [p] <[t] <[k] is only shown in M08’s data, not in F07’s data. F07 M08 p 0.216 <0.001 R-squared(%) 9.2 Not significant in F07’s content word set, but still in M08’s content word set.
Regression: word frequency In both speakers’ data, more frequent words tend to have shorter VOT, but the trends are not very significant. For both speakers, Buckeye frequency measure is slightly better than Celex frequency measure. Not significant in F07’s content word set, but still in M08’s content word set.
Regression: word frequency M08 Log Celex freq p R^2 (%) R^2 change of the model (%) 0.391 0.2 0 1.3 Log Celex freq p R^2 (%) R^2 change of the model (%) 0.169 0.3 9.2 9.4 Buckeye freq (speaker-specific) p R^2 (%) R^2 change of the model (%) 0.577 0 1.7 Buckeye freq (speaker-specific) p R^2 (%) R^2 change of the model (%) 0.067 0.7 9.2 9.6
Regression: phonetic context Two measures Category of the previous phone Coded as C(onsonant), V(owel), O(other sound), and N(on-linguistic) Category of the next phone
Regression: Phonetic context F07 M08 Previous phone category p R^2 (%) R^2 change of the model (%) 0.141 0.7 1.7 1.9 Previous phone category p R^2 (%) R^2 change of the model (%) 0.127 0.4 9.6 9.27 Next phone category p R^2 (%) R^2 change of the model (%) 0.563 0.4 1.7 0.6 Next phone category p R^2 (%) R^2 change of the model (%) 0.036 0.9 9.6 10.08
Regression: phonetic context VOT by previous phone category in F07 VOT by next phone category in M08
Regression: speech rate Three speed measures Duration of the next phone, in ms. Average speed of a 3-word period centered at the target word, measured in # of syll/s. Average speed of the pause-bounded stretch that contains the target word, measured in # of syll/s. All speed measures predict that words in faster speech tend to have shorter VOT
Regression: speech rate F07 M08 Duration of next phone p R^2 (%) R^2 change of the model (%) 0.014 3.2 1.9 5.1 Duration of next phone p R^2 (%) R^2 change of the model (%) 0.342 10.08 16.62 Average of the 3-wd stretch p R^2 (%) R^2 change of the model (%) <0.001 10.93 1.9 11.8 Average of the 3-wd stretch p R^2 (%) R^2 change of the model (%) <0.001 4.1 10.08 12.85 Average of the local stretch p R^2 (%) R^2 change of the model (%) <0.001 6 1.9 7.1 Average of the local stretch p R^2 (%) R^2 change of the model (%) 0.014 1.4 10.08 15.07
Regression: utterance position Utterance-final lengthening has been documented in the literature extensively. We code tokens for whether they are followed by silence. Number of tokens F07 M08 Non-final 146 312 final 9 34
Regression: utterance position F07 M08 non-final final non-final final
Regression: utterance position F07 M08 Utterance position p R^2 (%) R^2 change of the model (%) 0.021 2.8 11.8 19.11 Utterance position contributes to the variation in VOT Utterance position p R^2 (%) R^2 change of the model (%) 0.652 0.2 16.62 13.31 Utterance position doesn’t contribute to the variation in VOT
Regression: complete model F07 M08 Model performance Variable added R^2 (%) POA Buckeye Frequency 1.7 Previous phone category 1.9 Average speed of the 3-word stretch 11.8 Utterance position 19.11 Model performance Variable added R^2 (%) POA 9.2 Buckeye Frequency 9.6 Next phone category 10.08 Duration of the next phone 16.62
Regression: trends observed POA [p]<[t]<[k] Word class function words < content words Word frequency ??Higher frequency shorter VOT Here the word shorter or longer are used loosely, not referring to the specific effect of lengthening or shortening.
Regression: trends observed Phonetic category ??Preceded by vowel shorter VOT ??Followed by vowel longer VOT Speaking rate Faster speech shorter VOT Utterance position Utterance final longer VOT
Regression: trends observed Missing from the picture Contextual predictability Stress Disfluency Emotion
Overview Background Methodology Results Discussion Data Preliminary analysis Regression model Results Discussion
Discussion Individual differences Other between-subject factors Measurements Other between-subject factors Age Gender Average speaking rate
Discussion Relatively little variation is explained in the full model. (19.11% in F07 and 16.62% in M08) Factors missing from the picture: contextual predictability, stress, disfluency, etc. Limitation of linear regression model Non-linear effect Non-homogeneous effect Mixture of categorical and continuous variables
Discussion Echoing and challenging previous findings VOT and POA Canonical rule is observed in M08, but not in F07 Word frequency effect Overshadowed by word class distinction Utterance-final lengthening Significant in F07, but not M08 Speaking style? Content words vs. function words? Speed measures? Given speed measures, the partial correlation between vot and utterance-position in F07 is (1) if spd = next_dur, p = 0.0724 (2) if spd= spd_3wd, p = 0.4529 (3) if spd= str_spd, p =0.0743
Conclusion Still a long way to go to model VOT variation in spontaneous speech… Thanks! Any comments are welcome!
Thanks to Anonymous subjects Contributors to the Buckeye corpus Prof. Keith Johnson Members of the phonology lab in UC, Berkeley
Selected references Bell, A. et al. (1999) Forms of English function words - Effects of disfluencies, turn position, age and sex, and predictability. Proceedings of ICPhS-99 Gahl, S. In press. "Time" and "thyme" are not homophones: The effect of lemma frequency on word durations in a corpus of spontaneous speech. To appear in Language. Pitt, M. et al. (2005) The Buckeye Corpus of conversational speech: labeling conventions and a test of transcriber reliability. Speech Communication. Vol 45, pp: 90-95 Raymond et al. (2006) Word-internal /t,d/ deletion in spontaneous speech: Modeling the effects of extra-linguistic, lexical, and phonological factors. Yao, Y. (2007) Closure duration and VOT of word-initial voiceless plosives in English in spontaneous connected speech. UC Berkeley PhonLab report