Time-domain pitch extraction of speech signals using a digital filter

Abstract

An algorithm is described which derives pitch with respect to the beginning and end of individual pitch periods of speech signals in time domain using a nonrecursive digital filter. In a preliminary step, the higher formants are attenuated by a gap-function low-pass filter with a fixed first zero at about 1800 Hz. Analyzing the distribution of distances between adjacent zero crossings in this “pre-filtered” signal, the algorithm estimates the formant frequency F1. In the following step, a time-variant filter is applied whose zeros are situated around F1. The output signal of this filter shows a structure related to the waveform of the glottal voice source. Within a single pitch period, this “filtered signal” contains but one significant maximum which is marked a preliminary pitch period limitation. Since its position still depends on the structure of the original speech signal, the markers are shifted to the more reliable position of the first zero crossing which follows the significant maximum. To define the proper beginning and end of the pitch periods with correct phase, the markers are finally shifted by half a period of F1 against time. The algorithm is run on a digital computer and is being realized as a hardware device.