Speech Prosody 2004

Nara, Japan
March 23-26, 2004

Use of Tone Information in Continuous Cantonese Speech Recognition

Yao Qian (1), Tan Lee (1), Frank K. Soong (2)

(1) Department of Electronic Engineering, The Chinese University of Hong Kong
(2) Spoken Language Translation Labs, ATR, Kyoto, Japan

Cantonese, a syllabically paced, southern Chinese dialect, is also a tonal language where tones carry important lexical information. It is rich in tonal variations and each syllable can have up to 9 different tone patterns. In this paper we investigate how to incorporate the tone information into a large vocabulary continuous speech recognition system. A two-pass, post-processing scheme is proposed to utilize the recognized tones in rescoring the recognized N-best strings. Utterance level confidence measures of the N-best hypotheses are used in the rescoring process. It has been found from our experiments that weighted tone information can yield 8% relative improvement of the Chinese character error rate.

Full Paper

Bibliographic reference.  Qian, Yao / Lee, Tan / Soong, Frank K. (2004): "Use of tone information in continuous Cantonese speech recognition", In SP-2004, 587-590.