TR-I-0138 :1990.2

Alain de Cheveigne

Experiments in pitch extraction

Abstract:The aims of this study are threefold:

・Develop a methodology.

Many pitch extraction methods have been proposed. None is error-free, and it is often difficult to analyze the cause of failure. Error rate, as usually used for evaluation, counts the number of times the algorithm crosses the border between success and failure, and is therefore rather crude. It gives no indication of how close the algorithm comes to failure or success. The periodicity measure proposed here is more sensitive and offers better insights.

・Develop a parallel with auditory perception models.

The aim is to apply knowledge from time-domain models of auditory processing to the speech pitch extraction task. This is also an indirect test of the effectiveness of processing of such models.

・Develop a reliable speech pitch extraction method.

This is not the primary aim. Much research effort has already been invested in this problem (Hess 1983), and yet it has not been satisfactorily solved (e.g. Vaissiere 1989). It would be unrealistic to expect this particular study to succeed where so many others have failed.