Voiced/Unvoiced decision based on the bispectrum

We initiate an analysis of English phonemes using the bispectrum. The first part of our work demonstrates that fricative phonemes share a certain statistical property with Gaussian data not shared by purely voiced phonemes. Namely, approximations to the bispectra of fricative phonemes tend to zero, whereas, approximations to the bispectra of voiced phonemes do not. Indeed, the bispectra of voiced phonemes have complex structure as seen from their log-magnitude plots. A Fortran program based on the above principle processes speech filtered to 4 kHz and sampled at 8 kHz. Each 180 sample (22 ms.) block is declared to be voiced, unvoiced, or silent. In addition, a numerical measure of the mixed excitation is given. An advantage of this method over conventional ones is that it is less subject to noise interference.