Syllable-based large vocabulary continuous speech recognition

Abstract

Most large vocabulary continuous speech recognition (LVCSR) systems in the past decade have used a context-dependent (CD) phone as the fundamental acoustic unit. In this paper, we present one of the first robust LVCSR systems that uses a syllable-level acoustic unit for LVCSR on telephone-bandwidth speech. This effort is motivated by the inherent limitations in phone-based approaches-namely the lack of an easy and efficient way for modeling long-term temporal dependencies. A syllable unit spans a longer time frame, typically three phones, thereby offering a more parsimonious framework for modeling pronunciation variation in spontaneous speech. We present encouraging results which show that a syllable-based system exceeds the performance of a comparable triphone system both in terms of word error rate (WER) and complexity. The WER of the best syllable system reported here is 49.1% on a standard SWITCHBOARD evaluation, a small improvement over the triphone system. We also report results on a much smaller recognition task, OGI Alphadigits, which was used to validate some of the benefits syllables offer over triphones. The syllable-based system exceeds the performance of the triphone system by nearly 20%, an impressive accomplishment since the alphadigits application consists mostly of phone-level minimal pair distinctions.

Subword Modeling for Automatic Speech Recognition: Past, Present, and Emerging Approaches

“…Throughout the intervening years, however, a variety of alternative subword models have been studied in parallel, with the basic units including syllables [16], [17], acoustically defined units [18], [19], graphemes [20], and sub-phonetic features [21], [22], [23], [24], [25], [26]. Figure 1 serves as an informal summary of some of the main sub-word modeling approaches described in this article.…”

Modelling pronunciation variation with single-path and multi-path syllable models: Issues to consider

“…To alleviate the problems of the 'beads on a string' representation of speech, several authors propose modelling the spectral and temporal variation in speech 'implicitly' by using longerlength linguistic units as the basic building blocks of speech (Ganapathiraju et al, 2001;Hämäläinen et al, 2007a;Jones et al, 1997;Jouvet and Messina, 2004;Plannerer and Ruske, 1992;. For various reasons, most of these authors (Ganapathiraju et al, 2001;Hämäläinen et al, 2007a;Jones et al, 1997;Jouvet and Messina, 2004; suggest using syllable-length models.…”

“…Coarticulation effects, for instance, often stretch beyond the left and right neighbouring phones. The corresponding long-span spectral and temporal dependencies are not easy to capture with models that have as limited a window size as triphones (Ganapathiraju et al, 2001). Moreover, the pronunciation variants in the lexicon do not cover all variation in actual speech production (McAllaster and Gillick, 1999;.…”

“…For various reasons, most of these authors (Ganapathiraju et al, 2001;Hämäläinen et al, 2007a;Jones et al, 1997;Jouvet and Messina, 2004; suggest using syllable-length models. First, using syllables allows for a relatively compact representation of speech, while maintaining a manageable level of recogniser complexity.…”

“…First, syllable models with a sufficient amount of training data are used in combination with triphones (Ganapathiraju et al, 2001;Hämäläinen et al, 2007a;Jouvet and Messina, 2004;. In other words, triphones are backed off to when a given syllable does not occur frequently enough for reliable model parameter estimation.…”