Speaker-adaptive HMM-based speech recognition with a stochastic speaker classifier

A speaker-adaptive speech recognition method using a stochastic speaker classifier is proposed. The stochastic speaker classifier decides which spectral feature subspace is suitable for the input speaker by using integrated speaker Markov models. In the acoustic HMMs (hidden Markov models), the observation emission probabilities, are presented as joint probabilities for speaker individuality obtained from the speaker classifier and feature vectors, from the acoustic preprocessor. Evaluation experiments are performed using a telephone speech database of 50 command words and 10 Japanese digits. Using four integrated 9-state ergodic speaker hidden Markov models estimated from the command words uttered by 116 training speakers, the best word recognition accuracy of 98.1% is achieved for the 10 digits uttered by 116 test speakers. This is an improvement of 2% over the conventional pooled training method