First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

A New Japanese Text-To-Speech Synthesizer Based on COC Synthesis Method

Kazuo Hakoda, Shin-ya Nakajima, Tomohisa Hirokawa, Hideyuki Mizuno

NTT Human Interface Laboratories, Kanagawa, Japan

This paper describes a new Japanese text-to-speech synthesizer that produces far more natural and intelligible speech than existing synthesizers by using the new Context Oriented Clustering(COC) method. The COC method automatically generates speech unit variations from natural speech database. Preference tests show that the intelligibility of COC synthesized speech is better than that of the conventional dyad based method. A new LSP synthesizer which produces a wide frequency band of output speech is developed. The synthesizer is implemented with a general purpose Digital Signal Processor(DSP). Optimum design parameters, such as LSP order, parameter quantization bits are decided on the basis of spectral distortion and preference tests results. This synthesizer is constructed on a single PC board to permit easy installation in personal computers.

Full Paper

Bibliographic reference.  Hakoda, Kazuo / Nakajima, Shin-ya / Hirokawa, Tomohisa / Mizuno, Hideyuki (1990): "A new Japanese text-to-speech synthesizer based on COC synthesis method", In ICSLP-1990, 809-812.