Puretalk: a high quality Japanese text-to-speech system

This paper describes a high quality Japanese text to speech (TTS) system, PureTalk. This system is similar to the conventional diphone-based TTS using PSOLA except that PureTalk employs the following novel techniques which enable to produce more intelligible and natural-sounding speech: 1) two-stage duration modeling based on a linear regression technique, 2) F0 contour modeling using polynomial segment models, 3) sophisticated waveform unit selection, and 4) e cient waveform compression designed for TTS system. The result of the subjective hearing test shows that PureTalk achieves high quality under practical computation and memory requirement.