Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Incremental Speaker Adaptation Using Phonetically Balanced Training Sentences for Mandarin Syllable Recognition Based on Segmental Probability Models

Jia-lin Shen, Hsin-min Wang, Ren-yuan Lyu, Lin-shan Lee

Dept. of Electrical Engineering, Rm. 520, National Taiwan University Taipei, Taiwan

This paper presents a new incremental speaker adaptation technique for isolated Mandarin syllables using four sets of phonetically balanced sentences. This algorithm was based on a newly developed Segmental Probability Model(SPM) which was found specially suitable for isolated Mandarin syllable recognition. Each Mandarin syllable is conventionally divided into INITIAL and FINAL parts and based on the INITIAL/FINAL structure of Mandarin syllables, a segment sharing concept is first proposed. A computer algorithm was then developed to select automatically four sets of phonetically balanced sentences with different selection criterion from a large Chinese text corpus. After the four-stage adaptation procedure, the recognition rate for a new speaker can be improved from 63.05% to 92.96%.

Full Paper

Bibliographic reference.  Shen, Jia-lin / Wang, Hsin-min / Lyu, Ren-yuan / Lee, Lin-shan (1994): "Incremental speaker adaptation using phonetically balanced training sentences for Mandarin syllable recognition based on segmental probability models", In ICSLP-1994, 443-446.