Improved estimate of vocal-tract areas from the speech wave
Paul Milenkovic, Bishnu Saroop Atal · The Journal of the Acoustical Society of America · 1979
Areas obtained from linear prediction analysis of the speech wave often differ from actual voice-tract areas. Some of the differences are due to unrealistic assumptions made in the inversion procedure about the losses in the vocal tract, the spectrum of the glottal source, and the high-frequency spectrum of the vocal-tract impulse response. In this paper, we discuss the nature of errors introduced by these assumptions and describe a procedure for avoiding some of the errors. We discuss how lack of high-frequency spectral information creates ambiguities in mapping acoustic information to vocal-tract areas. Ambiguities are also caused by the uncertainity about the source characteristics. We resolve these ambiguities by requiring that the recovered vocal-tract shape be smooth; the smoothness requirement ensures that the dimensionality of the area function is not increased in the process of specifying the missing information. Results obtained from tests on both natural and synthetic speech signals will be presented.