Intelligibility modification of dysarthric speech using HMM-based adaptive synthesis system

Dysarthria is a manifestation of an inability to control and coordinate on one or more articulatory subsystems, which results in poorly articulated, slurred, and unintelligible speech. In order to enable a dysarthric speaker to communicate more efficiently with others, a text-to-speech synthesis system that generates speech in his voice, but without the errors he makes would be desirable. In this regard, the current work proposes a system, where the dysarthric speech is first recognized by an HMM-based speech recognition system. A sentence-level network is used to ensure 100% recognition accuracy. The recognized text is then synthesized by a speech synthesis system adapted to the dysarthric speaker's voice. This system replaces the sound units wrongly uttered by the dysarthric speaker, thereby improving intelligibility. The rate of synthesized speech is quite low for speakers with moderate and severe dysarthria. Therefore, the speech rate is modified using time-domain pitch synchronous overlap add (TD-PSOLA) technique. Degradation mean opinion score (DMOS) is used to prove that wrongly uttered sound units are replaced by correct sound units and that the synthetic speech is made more intelligible with the speaker's identity.