Predictor codebooks for speaker-independent speech recognition

Abstract

The authors examine the speech recognition capabilities of predictor codebooks under multi-speaker and speaker-independent conditions. Three structures of spectrum predictors, a forward predictor, a backward predictor, and an interpolator, are examined. Predictor codebooks are generated by the LBG algorithm with a small modification for predictor quantization. The predictor codebooks are then tested on a phone recognition task with three different measurements. The degradation in predictor-codebook performance was reduced by one-third under speaker-independent conditions. Finally, continuous-speech recognition experiments are carried out using the predictor codebook for multi-speaker and speaker-independent conditions. The results show that the backward-predictor codebook is very effective.<>