Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Speaker Adaptation Based on Transfer Vectors of Multiple Reference Speakers

Kazumi Ohkura, Hiroki Ohnishi, Masayuki Iida

SANYO Electric Co., Ltd., Hypermedia Research. Center, Speech Processing Laboratory, Osaka, Japan

This paper addresses a speaker adaptation method for speaker-independent phoneme HMMs (SI-HMMs). The type of these HMMs is a Gaussian continuous mixture density HMM. It is important for speaker adaptation to make up for a deficiency of training data. In our proposed method, the TRAnsfer VEctors of multiple Reference Speakers estimated from sufficient training data are used to make up for a deficiency of the training data. We call this method TRAVERSE. TRAVERSE was evaluated by 100 isolated words recognition experiment. The testing speakers were five males. The average recognition rate for the five speakers was 80% with SI-HMMs. Applying TRAVERSE to SI-HMMs, the recognition rates increased 86.1% and 90.5% with 1 and 5 isolated words for speaker adaptation.

Full Paper

Bibliographic reference.  Ohkura, Kazumi / Ohnishi, Hiroki / Iida, Masayuki (1994): "Speaker adaptation based on transfer vectors of multiple reference speakers", In ICSLP-1994, 455-458.