On the Use of Vector Quantization for Connected-Digit Recognition
Stephen C. Glinski · AT&T Technical Journal · 1985
Recent work at AT&T Bell Laboratories has demonstrated the efficacy of vector quantization in greatly reducing both the computational and memory requirements of isolated-word recognition systems. This efficiency is obtained at the expense of a marginal decrease in performance, and thus is an attractive approach. The purpose of this paper is to report on the results of a series of experiments in the application of vector-quantization strategies to a small-vocabulary, connected-word recognition task. Several strategies are investigated, including the use of speaker-trained code books versus universal code books, the use of binary and higher-order tree searches versus full searches of these code books, and the quantization of both test and reference frames versus reference frames only. For various strategies, the effect on error rate of varying the code-book size is also reported. Results indicate that the vector quantization approach is attractive for linear predictive coding-based connected-digit recognition.