Automatic measurement of voice onset time using discriminative structured prediction

Abstract

A discriminative large-margin algorithm for automatic measurement of voice onset time (VOT) is described, considered as a case of predicting structured output from speech. Manually labeled data are used to train a function that takes as input a speech segment of an arbitrary length containing a voiceless stop, and outputs its VOT. The function is explicitly trained to minimize the difference between predicted and manually measured VOT; it operates on a set of acoustic feature functions designed based on spectral and temporal cues used by human VOT annotators. The algorithm is applied to initial voiceless stops from four corpora, representing different types of speech. Using several evaluation methods, the algorithm's performance is near human intertranscriber reliability, and compares favorably with previous work. Furthermore, the algorithm's performance is minimally affected by training and testing on different corpora, and remains essentially constant as the amount of training data is reduced to 50-250 manually labeled examples, demonstrating the method's practical applicability to new datasets.

The private life of stops: VOT in a real-time corpus of spontaneous Glaswegian

“…This shows that, using our method, swift correction can easily be transferred to a new annotator, without introducing bias into the analysis. The overall performance of Auto-VOT on the Glaswegian corpus was good and similar to the results presented in Sonderegger and Keshet (2012), who carried out an evaluation of the performance of the algorithm on four different datasets comparing it to that of human transcribers for the same data. The results for the two datasets closest to our sample are the Switchboard corpus of American speech and the Big Brother UK dataset of spontaneous British speech; for both corpora, VOT detection windows were placed manually, rather than using force-aligned segment boundaries.…”

“…We developed a semi-automatic procedure to process large numbers of reliable VOT measures, using the Auto-VOT algorithm developed by Sonderegger and Keshet (2012). It was trained on an initial set of some 500 hand-measured tokens, applied to the force-aligned data, and then the algorithm's VOT predictions when applied to the full dataset were manually checked.…”

“…Here we considered VOT in a vernacular dialect where change in progress has been mooted (e.g., Johnston 1997), but for which only a few studies of read speech exist, and confounds of the influence of physiological age on VOT occur with those of possible inference of (apparent-time) change (Docherty et al 2011). We used a semi-automatic procedure based on Sonderegger and Keshet's (2012) Auto-VOT algorithm for predicting positive VOT, which requires some training data and the existence of boundaries roughly indicating the beginning of the stop. Using a fast manual coding scheme to correct the predictions of Auto-VOT, we gained reliable VOT measures for over 7,000 stops from 23 female speakers of Glaswegian stratified by age (decade of birth) and decade of recording.…”

“…We first automatically measured stop VOTs by applying the AutoVOT software (Keshet et al 2014), an implementation of the supervised learning algorithm described in Sonderegger and Keshet (2012). AutoVOT uses a set of handlabelled VOT measurements as a training set of stops to train a structured support-vector machine classifier.…”

Estimation of voice-onset time in continuous speech using temporal measures

“…Methods for the measurement of VOT fall into two categories: (a) those which explicitly identify the locations of the burst and voicing onsets through a set of customized acoustic-phonetic rules (knowledge-based), 4,6 and (b) those which train a learning machine (such as random forest, support vector machine) to estimate the VOT using some acoustic features corresponding to the stop-to-voiced-phone transition event. 8,9 Many of the high performing methods require phonetic transcription either to identify the segment of the speech signal containing the stop consonant through forced-alignment 4,9 or to focus the analysis on segments of the signal containing only one stop consonant. 8 Such methods are difficult to employ in a scenario where there is no transcription available.…”