Digital hardware for pitch detection

Abstract

A pitch detection algorithm, designed for robust performance and hardware simplicity, has been studied by simulation and has been implemented in digital hardware. The speech signal is low-pass filtered at about 900 Hz and sampled at a 10-kHz rate. A value of +1 is assigned to a speech sample if that sample value is above a positive threshold; likewise, a value of −1 is assigned if the sample is below a negative threshold. Otherwise, a value of 0 is assigned to the sample. The clipping thresholds are adapted to the speech level within a 30-msec analysis interval. The short-time autocorrelation function of this infinitely center clipped speech signal exhibits a strong peak at the pitch period for voiced speech and no peak for unvoiced speech. The peak location serves as the estimate of pitch period. A similar approach was proposed by Sondhi [IEEE Trans. Audio Electroacoust. (June 1968)]. Since only two data bits per sample are required to represent the infinitely center clipped speech, memory requirements are minimized and the computation of the autocorrelation function is reduced to a simple logical combination. The algorithm is easily implemented with standard TTL logic for real-time operation. The system performs well on bandpass speech (missing fundamental) and under conditions of low signal-to-noise ratio. The hardware was constructed for use in a speaker verification system, but can be easily adapted to any application where pitch detection is required.