A robust pitch boundary detector

An algorithm is presented which detects each and every pitch boundary of a voiced utterance. The algorithm is robust under additive white noise for up to 0-dB signal-to noise ratio. The algorithm is based on extraction of the peakedness feature and the similarity measure of the speech waveform. The authors model the peakedness of the properly rectified and filtered voiced speech envelope as a distorted sawtooth train. A matched filter is devised to extract the peakedness characteristics of the voiced speech and its periodicity. The impulse response of the matched filter is a sawtooth waveform with adjustable period length. The filter output achieves its peak value when the period of its impulse response matches the transient pitch period of the voiced speech