A Hybrid Acoustic-Articulatory Model of the Speech Spectrum
Jean Schoentgen, Abdellah Kacha, Francis Grenez · 2008
The object of the presentation is spectrum-to-area mapping in the framework of a speech analysis model. The proposal is to represent speech spectra via a hybrid model that comprises the morphologi-cal parameters of a quasi-lossless model of the main vocal tract cavity, which fixes the formant frequen-cies, and spectral parameters that specify the for-mant bandwidths and the formants and anti-formants owing to side-cavities as well as the glottal formant and spectral slope owing to the speech sound source. A study of feasibility shows that hybrid models are flexible enough to combine the ability of simulat-ing any spectral contour with the possibility to infer plausible tract shapes for a subset of phonetic seg-ments. 1