Effect of vector quantization on a continuous speech recognition system

Vector quantization techniques were applied to a continuous speech recognition system as a means of reducing both memory usage and computation time. The speech recognition system computes time-aligned distances between unknown speech segments and template frames. Vector quantization allowed the replacement of speech frames (vectors) with single index numbers which referenced an ordered set, or codebook, of representative frames. Two techniques for generating this codebook, clustering and covering, were examined. The covering technique provided a significant computational advantage over the clustering technique although both techniques generated codebooks which performed well in this task. Results are presented for a ten speaker, 100 word vocabulary experiment. Using speaker dependent codebooks, system performance levels were maintained while the number of distance calculations was reduced by a factor of 2 and the template storage required was reduced by a factor of 4.6. With an increase in error rate of about one third, these factors were 6.8 and 7.8 respectively.