Formant-based articulatory normalization and its application to vowel restoration

Abstract

Visual feedback of spontaneous speech is effective for articulatory training of deaf children and for speech rehabilitation of dysarthric patients. Especially, the visual representations of vowel formant frequencies have been used directly or indirectly for those purposes, because those acoustical parameters reflect the articulatory behavior. However, since not only the shape of the vocal tract but its size also affect the formant frequencies, minimization of the effect due to the differences in size is required. In such a speaker normalization, we defined a color space consisting of three circular ratios of formant frequencies and applied it to the color visualization of vowel sound. In this paper, we proposed a normalized articulation space as an expansion of the color space, where we assumed that neutral vowels of any speaker are mapped into a unique point. In addition, since the proposed articulatory space was regarded as the speaker independent representation of vocal tract shape, we also proposed a method to convert the modified articulation shape into the real formant space and applied it to the vowel restorations of disarthric speech.