A novel innovation based time domain pitch detector

We present a novel innovations based time-domain pitch detection technique for speech-like signals using one of our recursive least-squares ladder algorithms. The basic assumption is that the speech driving process consists of an approximately Gaussian part (unvoiced) and a jump part (voiced). The pitch pulse positions located by processing the innovations alone are known not to be very accurate due to phase-distortions, effects of zeros and inaccurate model parameter estimates. In our ladder form linear prediction recursions, a log-likelihood function is recursively computed (on-line) for each speech sample. The derivative of this log-likelihood function becomes a sensitive measure of extreme outliers of the speech waveforms, i.e., samples that very likely do not fit the Gaussian statistics. When combined with the innovations of our ladder algorithms a good statistic is obtained for locating the pitch pulses by thresholding. This pitch detection scheme has the advantage that all the necessary variables are already computed in the modeling ladder recursions and therefore is suited for fast on-line or even hardware (e.g., VLSI) implementations.