A robust pitch and voicing detector for harmonic coders

Pitch period doubling and tripling is a common problem in pitch detectors based on autocorrelation of the speech signal. In this paper we show the failure mechanism involved and present a new technique for detecting and correcting gross pitch errors which operates by detecting the resulting spectral nulls in the synthetic speech. This method is reliable and decreases the dependence on thresholds. We also present a pitch and voicing algorithm based on this approach which uses a local model of the speech signal. The algorithm is used in the Frequency Selective Harmonic Coder, one of eight candidates in the recent competition to select a new 2400 bps U.S. Government Voice Coding Standard.