A study of correlation methods for phoneme identification
A. Maynard Engebretson, John P. Tadlock · The Journal of the Acoustical Society of America · 1983
With current availability of high-speed digital signal processors and inexpensive memory, correlation methods of real-time speech recognition may be practical. In the methods presented here, speech is segmented into overlapping 25.6-ms windows. A spectrum is calculated for each window and correlated with speaker-independent reference patterns. Each windowed speech fragment is identified with the reference pattern yielding the highest correlation. Average spectral patterns are created automatically from “stable” segments of speech. A high correlation between spectra of adjacent windows is used as a measure of stability. The average patterns are categorized across speakers and averaged within categories to create the reference patterns. We have studied standard and Mellin (scale invarient) correlations in combination with linear magnitude and cepstral-smoothed log magnitude spectra. Mellin correlation normalizes spectra that are scaled with respect to frequency. Cepstral-smoothing reduces the number of coefficients required to represent the spectrum. Also cepstral coefficients can be normalized with regard to transmission characteristics. Results will be presented for vowels recorded from six male and six female speakers. [Work supported in part by NIH Grant NS-03856].