Improving pitch estimation with short duration speech samples

Hermes’ Sub Harmonic Summation (SHS) pitch determination algorithm is an effective technique for extracting the percept of pitch from human speech [1]. Effective determination of the pitch in a passage of speech is believed to be fundamental for higher level speech processing applications such as speech or speaker recognition. Of particular interest is the need to extract pitch from speech in less than ideal conditions eg. in the presence of noise or using very short analysis windows. In an attempt to deliver accurate pitch estimates from relatively short analysis windows this paper describes an evaluation of two forms of the SHS procedure: in one case, FFT-SHS, the procedure uses the conventional Fast Fourier Transform (FFT) in its spectral analysis step; in the second case, RAFT-SHS, the ReAssigned Fourier Transform (RAFT) technique [2] is used instead of the FFT.