Segment Specific Concatenation Cost for Syllable Based Bengali TTS

This paper proposes a new method of concatenation cost calculation for enhancing the optimality in unit selection. Instead of defining same set of concatenation costs for all types of speech unit transitions, costs are defined based on the type of unit transitions. Different types of unit transitions that can occur mainly in an utterance are voiced to voiced, voiced to unvoiced and unvoiced to unvoiced transitions. Natural measure of continuity is identified for each of these transitions, and costs are defined accordingly. For voiced to voiced transitions, in addition to spectral continuity, pitch and energy continuity metrics are proposed. In case of voiced to unvoiced and unvoiced to unvoiced transitions, silence duration embedded in the unvoiced region is proposed as the continuity metric. This approach of segment specific concatenation cost calculation improves the quality of syllable based text to speech synthesis. Listening tests provide a proof on the effectiveness of proposed methodology which has clearly shown the decrease in perceptual discontinuity at joins, and improvement in the overall quality of the synthesised speech.