The spectral autocorrelation applied to the linear prediction residual of speech for robust pitch detection

A robust method is described for the determination of pitch in voiced speech. The spectral autocorrelation function measures the regular harmonic spacing in the spectrum of the speech signal by applying an autocorrelation to the power spectral density. It is shown that this method is capable of accurate pitch tracking in additive noise. The preprocessing is critical for good performance, and in this case the speech spectrum is flattened by obtaining the residual from linear prediction analysis and the original speech. Application of suitable postprocessing enables this method to operate over a wide range of human pitch down to very low signal-to-noise ratios