Articulatory VCV Synthesis from EMA Data

This paper reports experiments in synthesizing VCV sequences with French unvoiced stop or fricative consonants, using a time-domain simulation of the vocal-tract system. The necessary dynamics of the vocal-tract shape are derived in two steps: first, time-varying parameters of an articulatory model are calculated automatically from electromagnetic articulography (EMA) data, using a method previously published by the first author; second, semi-automatic corrections are applied to properly account for consonantal events. Time-varying characteristics of the glottis are set using empirical rules. Friction noise is generated along the length of the whole vocal-tract instead of locally at the narrowest constriction. Spectrogams of satisfactory synthesis results are presented alongside those of real speech recorded simultaneously with the articulatory data. Corresponding audio files are available online.