Acoustic and EM waves for speech processing

Abstract

The use of low power EM sensors [Holzrichter, Burnett, Ng, and Lea, J. Acoust. Soc. Am. 103, 62 (1998)] enable the determination of several articulator boundary motions, simultaneously—e.g., glottis, jaw, tongue, palate, lips, etc.—leading to new speech processing algorithms. By comparing EM sensor methods of pitch detection to established cepstral and autocorrelation techniques, a tenfold increase in pitch accuracy, a 100-fold reduction in processing time, and robust performance in the presence of high background acoustic noise are obtained. Investigations of the details of the EM sensor signal from the human glottal region indicate that it can be used to provide information on the instant excitation function of voiced speech. Using the ARMA approach, the excitation information is removed from the acoustic signal, yielding well-defined transfer functions, formant locations, and pitch normalized data. In addition, using the glottal timing information, pitch synchronous processing has been employed using ARMA, LPC, Cepstral, and other approximation techniques. The excitation function and the corresponding filter coefficients for each speech unit (e.g., the a’s and b’s from the ARMA approximation) enable one to reconstruct good quality personalized speech, without the need for residual information.