XIMERA A new TTS from ATR based on corpus-based technologies.pdfVIP

XIMERA: A NEW TTS FROM ATR BASED ON CORPUS-BASED TECHNOLOGIES Hisashi Kawai?, Tomoki Toda??1, Jinfu Ni?, Minoru Tsuzaki?, and Keiichi Tokuda?? ?ATR Spoken Language Translation Research Laboratories (ATR-SLT) ?Graduate School of Engineering, Nagoya Institute of Technology {hisashi.kawai, jinfu.ni,minoru.tsuzaki}@atr.jp, {tomoki,tokuda}@ics.nitech.ac.jp 1 This work was conducted while Dr. Toda was working for ATR/NITECH. Dr. Toda is currently at Language Technologies Institute, Carnegie Mellon University. ABSTRACT This paper describes a new concatenative TTS system under development at ATR. The system, named XIMERA, is based on corpus-based technologies, as was the case for the preceding TTS systems from ATR, namely ν-talk and CHATR. The prominent features of XIMERA are (1) large corpora (a 110- hours corpus of a Japanese male, a 60-hours corpus of a Japanese female, and a 20-hours corpus of a Chinese female), (2) HMM-based generation of prosodic parameters, and (3) a cost function for segment selection optimized based on perceptual experiments. A perception test that evaluated the naturalness of synthetic speech for XIMERA and 10 TTS products, including CHATR, showed that XIMERA outperformed the other ten. 1. INTRODUCTION Corpus-based technologies are undoubtedly a major trend in contemporary TTS (Text-To-Speech) systems. Among them, waveform concatenation techniques are widely adopted in commercial and experimental TTS systems for their natural- sounding output speech. These seem to be the most promising technique for achieving a practical TTS system at present. ATR, as one of the pioneers in corpus-based speech synthesis technology, has made contributions to the progress of the technology through various studies, which lead us to the development of two TTS systems, namely ν-talk [1] and CHATR [2,3]. The authors are currently developing a third TTS system named XIMERA (pronounced lik