Vocal tract areas from LPC—A comparison of several techniques

Abstract

The need for good initial estimates of cross‐sectional vocal tract areas [e.g., Schroeter, Larar, and Sondhi, J. Acoust. Soc. Am. Suppl. 1 80, S19 (1986)] has renewed our interest in LPC techniques. Starting with speech synthesized from a vocal tract model with known cross‐sectional areas [Schroeter and Sondhi, J. Acoust. Soc. Am. Suppl. 1 78, S6 (1985)], LPC‐derived areas that closely approximate the areas used to control the synthesizer are sought. In a comparative study, the LPC method used to derive the predictor polynomial (autocorrelation, covariance, and stabilized covariance methods) as well as its particular application with respect to location of pitch epochs is varied; pitch synchronous and asynchronous analyses are performed with different assumptions regarding the glottal termination. Spectral shaping and source components in the inverse filter are removed and formant shifts due to yielding wall losses are considered. Areas are computed using the standard Wakita recursion [Wakita, IEEE Trans. ASSP‐27, 281–285 (1979)] proceeding from the glottis to the lips with the first supraglottal section set to the known value (used for synthesis). Results are presented for steady vowels and for voiced segments produced with a time varying vocal tract.