Computer Recognition of Speech Utilizing Zero-Crossing Information.
NAVAL POSTGRADUATE SCHOOL MONTEREY CALIF
Pagination or Media Count:
The nature of speech sounds is studied with particular emphasis on the information bearing elements of speech. The association of the amplitude clipped-speech zero-crossing rate, formant frequencies and information content of a speech signal is presented and capitalized upon to produce readily extractable first and second formants from the speech wave. Various methods of processing the formants to generate unique patterns for particular sounds are attempted, with a time plot of the arithmetic difference of the two formants being explored in detail. The object being to obtain machine recognition of speech. Computer machine language programs are prepared to realize an Euclidean comparison of spoken numbers zero to nine against a previously stored dictionary. Testing showed this type processing satisfactory for some voices, but not readily extendible to many voices with the same dictionary. Methods of overcoming this shortcoming are suggested. Author
- Voice Communications