Approach to vowel normalization
Hisashi Wakita · The Journal of the Acoustical Society of America · 1975
This study is to test the hypothesis that perceptually equivalent vowels which are produced by various speakers have a statistically unique vocal-tract area function. The test was made in the frequency domain. The vocal-tract area function was computed from a measured set of vocal-tract resonance frequencies along with the tract length. Then the area function was normalized to one having a certain reference length, from which normalized resonance frequencies were computed. The study was made with ten vowels in the context of “hVd” produced by seven male and six female speakers. Both unnormalized and normalized distributions of F1−F2, F1−F3, and F2−F3 were obtained. It was ascertained that the normalized distributions are substantially more compact than the unnormalized ones. This result is particularly useful for interspeaker normalization in automatic speech recognition. [This work was supported by the Office of Naval Research.]