An improved SIFT method for pitch estimation of speech

This paper presents an improved SIFT (Simplified Inverse Filtering Technique) method for accurate pith estimation. In order to save computing time as well as ensuring the precision of autocorrelation, different re-sampling ratios are utilized during the process of LPC (Linear Predictive Coding) coefficients analysis and inverse filtering respectively; Furthermore, for the sake to satisfy the range and accuracy of pitch frequency simultaneously, Hamming-weighting is adopted for searching the reliable peak value on the autocorrelation curves, and a four-point non-linear pitch-smoothing algorithm is designed to avoid incoherent errors for an example in transient speech frames. Finally, the smoothed pitch contour is extracted and time-normalized pitch frequencies are calculated, which can then be used as the feature of a speech utterance in speech recognition or speaker recognition systems. Further Experiments show that the present method for pitch estimation of speech has good performance.