An inverse filtering technique for efficient vowel classification

Abstract

Classification of steady‐state vowel sounds by formant frequency clustering is well documented in the literature [G. E. Petersen and H. L. Barney, J. Acoust. Soc. Am. 24, 175–184 (1952)]. A more efficient and equally effective scheme for classifying steady‐state vowels has been developed using partial correlation (PARCOR) coefficients that arise as intermediate parameters in a least squares adaptive lattice inverse filter. A sixth‐order autoregressive time series was synthesized from three formant/bandwidth pairs for each vowel utterance of the Petersen and Barney data. The PARCOR coefficients of the inverse filtered processes exhibited the same graphical clustering behavior as the formant frequency data, and numerical distance measures showed the two sets of parameters to be equivalent. Because the PARCOR coefficients are used directly to classify vowels and because a pth order lattice filter simultaneously generates PARCOR coefficients for all lesser order filters, PARCOR coefficients provide an efficient, minimal parametrization. These results motivate the study of this technique for processing and classifying more complex and time‐varying speech signals. [Work supported by Naval Sea Systems Command.]