Normalization for vocal tract differences using long term average spectrum.
A. Davi Vitela, Andrew J. Lotto, Brad H. Story · The Journal of the Acoustical Society of America · 2009
Differences in vocal tract size and anatomy lead to substantial differences in acoustic realization of vowels. Much of the variability in vocal tract area functions is captured in the area function of the “neutral” (non-articulating) vocal tract of each talker. A listener could perceptually normalize talker variability substantially by extracting the neutral vocal tract and using it as a referent. Alternatively, the listener could extract the auditory representation of the output of the neutral tract and use it as a referent for the auditory representation of the vowel space. We will present such a model of perceptual talker normalization for vowels and demonstrate its effectiveness for a variety of vocal tract shapes and lengths. Additionally, we will provide evidence that this normalization may be implemented in an auditory process that (i) extracts the long term average spectrum from running speech (providing an estimate of the neutral vowel), (ii) represents new input relative to this spectral average (providing normalization), and (iii) is not specific to speech. Additional analyses and perceptual results will demonstrate how this model can extend to normalization of consonant perception. [Work supported by NIH/NIDCD to A.J.L. and B.H.S. and Diversity Supplement to A.D.V.]