HMMs for Acoustic Modeling: Beyond the Problem of Local Stationarity

Automatic speech recognition (ASR) systems based on hidden Markov models (HMMs) are effective under many circumstances, but do suffer from some major limitations that limit their applicability in real-world environments. In particular, the 'local stationarity' requirement, implicit in standard HMMs, appears to be constraining. This paper reviews the major concepts underlying HMMs for ASR, pointing out a number of alternatives to enhance the basic paradigm in order to overcome the local-stationarity problem. Alternatives considered herein are: (i) introduction of dynamic acoustic features; (ii) definition of contest-dependent acoustic units; (iii) integration of segmental information; (iv) combination of artificial neural networks within hybrid HMM/connectionist architectures