Segmentation of the speech signal based on changes in energy distribution in the spectrum

A simple algorithm is proposed for automatic phonetic segmentation of the acoustic speech signal on the MERA 303 desk-top minicomputer. The algorithm is verified with Polish linguistic material spoken by two subjects. The proposed algorithm detects approximately 80 percent of the boundaries between enunciated segments correctly, a result no worse than that obtained using more complex methods. Speech recognition programs are discussed as speech perception models, and the nature of categorical perception of human speech sounds is examined.