Integrating Recurrent Neural Networks and Dynamic Programming in an Hybrid Speech Recognition System: Concepts and Preliminary Results

This report introduces an hybrid speech recognition system for Speaker Independent (SI), continuous speech with a small vocabulary (sequences of Italian digits). The hybrid is based on parallel, state-space recurrent neural networks trained to perform a-posteriori state probability estimates of an underlying hidden Markov model (HMM) with fixed transition probabilities, given the sequence of acoustic observations. Training is accomplished in a supervised manner, relying on a prior Viterbi segmentation. Decoding is accomplished in a Dynamic programming framework, i.e. transitions along paths of the HMM, realizing a Viterbi decoding criterion. Preliminary experimental results of state-emission probability estimation and recognition of noisy signals acquired on the telephone line are presente