Automatic phoneme recognition with Segmental Hidden Markov Models

A new technique is presented for the joint recognition and segmentation task formulated for a speaker independent continuous phoneme recognition and segmentation system. We investigate a strictly probabilistic approach for simultaneous phoneme sequence segmentation and recognition. The implemented automatic phoneme recognition system integrates phoneme length statistics as well as phoneme transition statistics into the Segmental Hidden Markov Model (SHMM). A variation of the Viterbi Search algorithm is employed for estimating the most likely sequence of phonetic symbols as well as their corresponding segment boundaries. The Segmental HMM topology essentially models a phonetic symbol string with a double layer Hidden Markov Model (HMM), with each phonetic symbol in the Segmental HMM modeled with a left-to-right HMM. Our approach lays the groundwork for further expansion of Segmental HMM design to context dependent continuous phoneme recognition systems.