A Hybrid Stochastic Connectionist Approach to Automatic Speech Recognition

This report focuses on a hybrid approach, including stochastic and connectionist methods , for continuous speech recognition. Hidden Markov Models (HMMs) are a popular stochastic approach used for continuous speech, well suited to cope with the high variability found in natural utterances. On the other hand, artiicial neural networks (NNs) have shown high classiication power for short speech utterances. Therefore, we have built a hybrid system with the advantage of both Hidden Markov Models and Neural Networks. The basic idea is as follows: build a codebook from the Time-Delay Neural Networks (TDNN) output units and train HMMs using the Fuzzy-VQ algorithm. We trained several discrete HMMs for the recognition task of the Japanese phonemes using just one TDNN-generated codebook. We achieved a recognition rate of 96.1%, and in so doing, increased the recognition rate of the discrete HMMs by 7.1%. The results are an obvious proof of the possible collaboration of two diierent systems aimed at the same task.