Continuous phonetic speech recognition

Abstract

We report on research to develop an automatic phonetic recognition system for continuous speech. The system is based on a hidden Markov model (HMM) representation of phonemes in context. That is, the model parameters depend on the left and fight phonetic contexts for each phoneme. The HMM structure is the same for all phonemes, but the model parameter values are different for different phonemes and different contexts. Automatic training procedures are used to adjust the model parameters, using a given set of training speech data. A major focus of our work is to maximize the robustness of the phonetic models given a limited set of training data. An initial system is now operational. Results of phonetic recognition accuracy in continuous speech will be presented. [Work supported by ARPA and monitored by ONR.]