Abstract
A real-time speaker-dependent connected word recognition system with speaker adaptation is described. Cepstral analysis is performed using a digital signal processor. The system generates word candidates by continuous dynamic time warping (DTW) between input speech and word templates. With DTW-LSI chips, the matching between an input pattern and 500 word templates can be carried out in real time. Sentence syntax rules are represented by a finite-state network model, and a real-time search is performed using parallel processors. The system can handle an 8000-word (maximum) vocabulary task, if focus control is performed and the words to be recognized simultaneously are in groups of less than 500. The system has a speaker adaptation function based on vector quantization. The recognition system has been evaluated in a 191-word vocabulary expert system for a nuclear power plant and achieves sentence-recognition accuracies of 94.7% and of 93% with the speaker adaptation. The system can realize a speech dialogue expert system for practical use.<>