Improved HMM phone and triphone models for real-time ASR telephony applications

The development of human-machine dialog applications for messaging and information retrieval over the telephone poses stringent requirements on the accuracy and speed of the automatic speech recognition (ASR) system. The authors describe strategies for improved acoustic-phone modeling directed towards increasing recognition accuracy while keeping the number of phone units low. Specifically, the paper considers: (1) the development of an improved set of head-tail context-dependent (CD) triphones; (2) a novel criterion for better selection of the number of states assigned to each phone unit based on the coefficient of variation measure of feature components in HMM-Gaussians. The performance of the models is evaluated using data that represent real telephony applications.