Harmonic peaks method for voice separation

Abstract

This paper addresses the problem of recognizing a target voice when it is corrupted by a co-channel interfering voice. First, the F0 contour of the target voice is robustly extracted by using the revised highest likely common fundamental algorithm proposed by Screenivas. By using this contour, the harmonic peaks of the target voice are extracted. The harmonic peaks carry the information of the formants of a vowel and can be used as a front-end feature in a speech recognizer. Moreover, the harmonic peaks of the target voice are changed little even in the presence of an interfering voice. A recognizer aimed at recognizing the Mandarin finals was developed, based on the harmonic peaks method as well as the conventional LPC cepstral coefficients method. By comparing the results of these two methods, the harmonic peaks method shows better performance in the presence of a co-channel interfering voice.