Recent progress in articulatory modeling for speech recognition

Abstract

The automatic speech recognition research community has experimented with models of speech articulation for several decades, but such models have not yet made it into mainstream recognition systems. The difficulties of adopting articulatory models include their relative complexity and dearth of data, compared to traditional phone-based models and data. This talk will review the current state of articulatory models and will describe one particular approach to incorporating such models in modern speech recognition. In this approach, the articulatory variables are based on the vocal tract variables of articulatory phonology, and the models are represented using dynamic graphical models, a generalization of the more commonly used hidden Markov models. This approach allows the probabilistic modeling of asynchrony between articulators and reduction in articulatory gestures. Results will be presented showing improvements in lexical access using this type of articulatory model with automatically learned context-dependent articulatory feature distributions. Recent efforts to mitigate the data sparseness problem, including manual and automatic transcription, will also be presented.