Tone features for speech recognition

Abstract

Robust acoustic tone features are achieved first by the introduction of on-line, look-ahead trace back of the fundamental frequency (F0) contour with adaptive pruning, this fundamental frequency serves as the signal preprocessing front-end. The F0 contour is subsequently decomposed into lexical tone effect, phrase intonation effect, and random effect by means of time-variant, weighted moving average (MA) filter in conjunction with weighted (placing more emphasis on vowels) least squares of the F0 contour. The phrase intonation effect is defined as the long-term tendency of the voiced F0 contour, which can be approximated by a weighted-moving average of the F0 contour, with weights related to the degree of the periodicity of the signal. Since it is irrelevant from lexical tone effect, therefore it is removed by subtraction of the F0 contour under superposition assumption. The acoustic tone features are defined as two parts. First is the coefficients of the second order weighted regression of the de-intonation of the F0 contour over neighbouring frames, with window size related to the average length of a syllable and weights corresponding to the degree of the periodicity of the signal. The second part deals with the degree of the periodicity of the signal, which are the coefficients of the second order regression of the auto-correlation, with lag corresponding to the reciprocal of the pitch estimate from look-ahead tracing back procedure. These weights of the second order weighted regression of the de-intonation of the F0 contour are designed to emphasize/de-emphasize the voiced/unvoiced segments of the pitch contour in order to preserve the voiced pitch contour for the semi-voiced consonants. The advantage of this mechanism is, even if the speech segmentation has slightly errors, these weights with look-ahead adaptive-pruning trace back of the F0 contour served as the on-line signal pre-processing front-end, will preserve the pitch contour of the vowels for the pitch contour of the consonants. This vowel-preserving property of the tone features has the ability to prevent model parameters from bias estimation due to speech segmentation errors.