A neural network controlled adaptive search strategy for HMM-based speech recognition

A novel adaptive search method for an HMM (hidden Markov model-based continuous speech recognition system is described. A speech recognition system usually uses a heuristic search technique such as a beam search technique requiring a large number of phoneme verifications to achieve optimal search. To reduce this verification overhead and speed up the recognition process, a trainable adaptive search algorithm controlled by a neural network using observable features is introduced. This framework has the potential to improve automatically and dynamically the search mechanism by a neural network training procedure. Experimental comparisons with conventional beam search techniques show that the algorithm is effective in reducing the number of phoneme verifications with little degradation in recognition performance.<>