ABSTRACT
According to Fourier analysis, any periodic function can be analyzed as an infinite series of trigonometric functions (sets of sines and cosines). The kernel of decay cosine yields an extension for the previous frequency-based, sieve-type detection algorithm by giving smooth peaks for decaying amplitudes with the harmonics of the signal correlation. The sequential outline of the RAPT algorithm is: 1) Providing speech samples with their sampling rate and with a reduced sampling rate. 2) Periodically, computing normalized cross-correlation function (NCCF) of the reduced sampling rate speech signal with lags in the F0 range. 3) Indicating the locations of maximum at the 1st pass of NCCF. 4) For the vicinity of the peaks in that 1st pass, calculate the NCCF for the original sampling rate. 5) Again, finding the maximum in that NCCF. Obtaining the location and amplitude of the modified peak. 6) For each peak obtained from the NCCF (high resolution), estimate the F0 of the processed frame. 7) The hypothesis of the frame for unvoiced/voiced is advanced for each frame. 8) Finding the group of the NCCF peaks via optimization process for the unvoiced/voiced hypotheses for all the frames which have the best match with the above characteristics. 9) Using the well-known speech pitch tracking algorithm (PTA), RAPT has the following differences: - PTA computes the NCCF in the linear prediction coding (LPC).