Speaker Normalization for Automatic Word Recognition

Abstract

Automatic word-recognition systems typically exhibit marked degradation in performance when used on speakers other than those who contributed the “learn” data. Extension of word-recognition systems to accommodate new speakers usually involves extensive retraining. A procedure which would measure and “adapt” to characteristics of a new speaker would allow for improved recognition rates without retraining. One such procedure has been to measure maxima and minima of formant frequencies and adjust the recognition accordingly [L. J. Gerstman, “Classification of Self-Normalized Vowels,” AFCRL Conference on Speech Communication and Processing Reprints (1967)]. Of particular significance are procedures which involve little or no manual intervention and allow the recognition to remain “automatic.” This paper reports on several approaches, including use of the sample covariance matrix as a purely statistical method, and use of a “phoneme-map” as a more subjective manual intervention method. The various approaches are discussed in terms of degree of recognition improvement, amount of additional complexity, and amount of nonautomatic effort required.