Information conveyed by vowels about other vowels
Hector Raul Javkin, Elaine Drom, Carol Christie, Gaston R. Cangiano · The Journal of the Acoustical Society of America · 2004
Rapid adaptation to different speakers has become an important issue in speech recognition (ASR) but has been known in human listeners for a long time. Ladefoged and Broadbent [J. Acoust. Soc. Am. 29, 98–104 (1957)] demonstrated that human perception of synthesized vowels occurring in monosyllabic words (bit, bet, bat, but) can be changed by changing the formants of an introductory phrase (Please say what this word is). These stimuli meant that the introductory phrase ranged over the same portion of the vowel space (front vowels, or relatively high F2), thus facilitating listeners adaptation. To further test the limits of human adaptation, we replicated the experiment keeping the same words, substituting introductory phrases consisting of back (low F2) vowels, and maintaining a similar level of low-quality synthesized speech. The effects are difficult to replicate with natural or high-quality synthetic speech. However, we will suggest that low quality speech is analogous to the low-dimensionality representation of speech of many ASR front ends, which discard, for example, information as to the higher formants. Therefore, these findings are relevant to the use of adaptation in improving ASR. [Work supported by a Faculty Small Grant from San Jose State University.]