Towards non-uniform unit HMMs for speech recognition

Abstract

A novel acoustic modeling algorithm that generates non-uniform unit HMMs to effectively cope with spectral variations in fluent speech is proposed. The algorithm is devised for the automatic iterative generation of long-span units for the non-uniform modeling. This generation algorithm is based on an entropy reduction criterion using text data and a maximum likelihood criterion using speech data. The effectiveness of the non-uniform models was confirmed by comparing likelihood values between the long-span unit HMMs and the conventional phoneme-unit HMMs. Preliminary results suggest that non-uniform unit HMMs achieve higher performance than phoneme-unit HMMs.<>