Modeling microsegments of stop consonants in a hidden Markov model based word recognizer

Abstract

The motivation of this study is the poor performance of speech recognizers on the stop consonants. To overcome this weakness, word initial and word final stop consonants are modeled at a subphonemic (microsegmental) level. Each stop consonant is segmented into a few relatively stationary microsegments: silence, voice bar, burst, and aspiration. Microsegments of certain phonemically different stops are trained together due to their similar spectral properties. Microsegmental models of burst and aspiration are conditioned on the adjacent vowel category: front versus nonfront vowels. The resulting context-dependent microsegmental hidden Markov models (HMMs) for six stops possess the desired properties for a compromise between modeling accuracy and modeling robustness. They allow the recognizer to focus discrimination onto those regions of a stop that serve to distinguish it from other stops. Use of these models in recognition experiments for word lists consisting of CVC words reduces the error rate by 35% compared with the result obtained by using one HMM for each stop phoneme.

Speech analysis and recognition using interval statistics generated from a composite auditory model

“…All the syllables are uttered with a short intervening pause by native English speakers in a normal office environment. 6 This task was chosen because of its speaker-dependent nature, which complements the speaker-independent results described in Experiment I, and also because stop confusion is a particularly difficult problem identified from earlier studies [26].…”

The automatic recognition of stop consonants using hidden Markov models

“…In our opinion, one of the major reasons for this is that all of the systems, except for [6], referred to, do not take into account that stops consist of several well defined segments. This view is shared by Deng et al [6] who, by modelling the segments of the stop with hidden Markov models, found that the error rate of their speaker dependent large vocabulary recognition system was reduced by 35%, compared to the system that utilized a single Markov model for each phoneme.…”

“…Each of the segments described may be incorporated into a hidden Markov model. The works by Deng [6] and Levinson [9] give references on the subject and the paper by Rabiner and Juang [lo] is an excellent tutorial on hidden Markov models. In this study the following segments have been defined for the stop.…”

Speaker-independent consonant classification in continuous speech with distinctive features and neural networks

“…Through the years, research has been devoted to apply such knowledge to phonetic decoders based on artificial intelligence methods (e.g., De Mori et al, 1987;Lamel, 1993 ), to decoders based on statistical models (e.g., hidden Markov models, as in Shwartz et al, 1985;Lee, 1989;Deng et al, 1990; Bartkova and Jouvet, 1991;Deng and Erler, 1992). For example, in the model of Deng and Erler, 1992 the hidden states in the Markov models represent bundles of distinctive features.…”