An articulatory-feature-based HMM for automated speech recognition

Abstract

Traditional HMM-based speech recognizers model the acoustics as a set of disjoint segmental units. Little meaning is attached to internal states of the model and there are difficulties in representing context-dependency effects observed in the acoustic waveform. In this work, an articulatory-feature model is defined as an HMM in which each state represents a point in a space of quantized articulator configurations. A single large ergodic model then contains all feasible configurations and is sufficient to represent any vocabulary. Each word in the vocabulary is described by a sequence of target articulatory configurations, currently derived from the phonetic transcription. Context dependency is explicitly handled by those states visited between articulatory-target states. Each model state has some meaning in terms of the physical state of the vocal production process, so linguistic and physiological knowledge can be incorporated to restrict the model evolution and improve recognition performance. The development of the quantized articulatory space, target articulatory-feature sequences, and feature-evolution constraints is described for a large-vocabulary word recognition system. Recognition results show it to be competitive with the traditional phoneme model. [Work supported by NSERC, Canada and ITRC, Ontario.]