Evaluation of three pitch tracking algorithms at several signal-to-noise ratios.
Source: Acustica 83(3):567-71.
Abstract: The quality of pitch extraction is crucial in feature extracting hearing aids that provide optimum pitch information to the profoundly hearing impaired listener. Three pitch tracking algorithms were studied for their performance as a function of signal-to-noise ratio. The cross-correlation of the time signal followed by dynamic programming (CCF-DP) as implemented in the formant-extraction program of Entropic was the first algorithm studied. The second one was the sub harmonic summation (SHS). The third and fourth algorithms were from an artificial neural net consisting of a multi-layer perceptron (MLP; 3). Maximum accuracy in pitch estimation was seen in the CCF-DP Algorithm. The pitch estimation with SHS and MLP was nearly equal when the effective time window of MLP was increased to 40 ms. The most robust voicing classification was seen for the MLP, next for the SHS and the least for. At the signal-to-noise ratios of 0 and -5 dB S/M , the classification by the CCF-DP was the least robust. (CIRRIE Abstract)
Institution: Lab. of Exp. Audiology, Univ. Hospital Utrecht, Netherlands