Improved hidden Markov model adaptation method for reduced frame rate speech recognition

An improved hidden Markov model (HMM) adaptation method is proposed for the recognition of reduced frame rate speech. In previous studies of this kind of model adaptation, individual models were first adapted and then the adapted models were connected to form the word network for a speech recognition system. This adapting-then-connecting approach can produce transitions that skip too many states and increase insertion errors. In the proposed new method, the transition probabilities from the states of a HMM to the states of a following model are calculated by first connecting the two models to form a combined model and then adapting the transition probability matrix of the combined model. This connecting-then-adapting approach can rectify the problem of skipping too many states, and therefore reduce insertion errors. Experiments were conducted to investigate the effectiveness of the proposed method and the experimental results show that the proposed new method can obtain better recognition accuracy.