Automatic estimation of voice onset time for word-initial stops by applying random forest to onset detection

Abstract

The voice onset time (VOT) of a stop consonant is the interval between its burst onset and voicing onset. Among a variety of research topics on VOT, one that has been studied for years is how VOTs are efficiently measured. Manual annotation is a feasible way, but it becomes a time-consuming task when the corpus size is large. This paper proposes an automatic VOT estimation method based on an onset detection algorithm. At first, a forced alignment is applied to identify the locations of stop consonants. Then a random forest based onset detector searches each stop segment for its burst and voicing onsets to estimate a VOT. The proposed onset detection can detect the onsets in an efficient and accurate manner with only a small amount of training data. The evaluation data extracted from the TIMIT corpus were 2344 words with a word-initial stop. The experimental results showed that 83.4% of the estimations deviate less than 10 ms from their manually labeled values, and 96.5% of the estimations deviate by less than 20 ms. Some factors that influence the proposed estimation method, such as place of articulation, voicing of a stop consonant, and quality of succeeding vowel, were also investigated.

Automatic measurement of voice onset time using discriminative structured prediction

“…The algorithm's performance can then be reported either as the full (empirical) CDF of automatic/manual Lin and Wang, 2011). Reporting statistics about the CDF of automatic/manual differences is a standard evaluation method in ASR tasks, such as forced alignment of phoneme sequences, where the goal is to predict the location of boundaries in a speech segment (e.g., Brugnara et al, 1993;Keshet et al, 2007).…”

“…Previous work has used automatic measurements for speech recognition tasks Ramesh, 1998, 2003;Ali, 1999;Stouten and van Hamme, 2009), phonetic measurement (Fowler et al, 2008;Tauberer, 2010), and accented speech detection (Kazemzadeh et al, 2006;Hansen et al, 2010). Some studies, like ours, focus largely on the problem of VOT measurement itself, and evaluate the proposed algorithm by comparing automatic and manual measurements (Stouten and van Hamme, 2009;Yao, 2009;Hansen et al, 2010;Lin and Wang, 2011). Our approach differs from all previous studies except one (Lin and Wang, 2011) in an important aspect.…”

“…Some studies, like ours, focus largely on the problem of VOT measurement itself, and evaluate the proposed algorithm by comparing automatic and manual measurements (Stouten and van Hamme, 2009;Yao, 2009;Hansen et al, 2010;Lin and Wang, 2011). Our approach differs from all previous studies except one (Lin and Wang, 2011) in an important aspect. Instead of using a set of customized rules to estimate VOT, our system learns to estimate VOT from training data.…”

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

“…and Van Hamme (RS), 5 the random-forest-based method by Lin and Wang (RF) 9 and structured-prediction-based method by Sonderegger and Keshet (SP). 8 All of these studies report results on the TIMIT database using the same validation criterion.…”

“…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.…”