On the Application of Vector Quantization and Hidden Markov Models to Speaker-Independent, Isolated Word Recognition

There currently exist two standard approaches to isolated word recognition, namely, feature extraction methods and statistical pattern recognition models. A statistical pattern recognition approach has the property of being a nonparametric approach to recognition and therefore is widely used in most commercial and industrial recognizers. 1-6 The feature-based approach to recognition has been primarily used in the (computationally) less expensive systems, and as a basis for recognition of continuous speech (in conjunction with segmentation and labeling algorithms). 4-9 In the past few years a new approach to speech processing has been proposed, namely, using probabilistic functions of Markov models. This approach has been applied at the Institute for Defense Analyses for speaker recognition,10 and at Carnegie Mellon University and IBM to solve problems in continuous speech recognition 1112 with good success. Based on its success in these related areas of speech processing, a question that arises naturally is how well these probabilistic models would work on problems in isolated word recognition. It is the prime purpose of this paper to provide an answer to the question posed above. Before discussing the approach we have taken to get at the answer, we must first describe the structure of a word recognition system based on (hidden) Markov models (HMM). As in most recognition systems we assume we have a labeled training set of data from which we build a series of Markov models, one for each vocabulary word.