Speech processor system for pitch period extraction using prediction

Abstract

A computational algorithm and an implementation thereof is described herein for determining the pitch period of voiced speech in real time. All processing is performed in the time domain employing the prediction residual or error signal of a 10th-order Itakura cascade adaptive linear predictor or filter as the input signal. The output (pitch period) of the algorithm and the implementation thereof is updated each sample period based on analysis of the present and past input samples. Pitch period is determined by locating the sharp pitch peaks in the short term power of the prediction residual. The instantaneous pitch period is the time separation of two adjacent pitch peaks. The algorithm and implementation thereof employs a time moving search window and a time varying threshold level to locate pitch peaks. Various tests and procedures are incorporated into the algorithm and the implementation thereof to handle the special cases of false and missed pitch peaks. Detected errors are corrected within the algorithm and the implementation thereof by utilizing past data. Unlike the correlation or averaging methods of pitch extraction which require large amounts of storage and arithmetic operations, the time domain method of this invention requires a minimal amount of storage and only simple comparisons of amplitudes.