Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

A Speaker Adaptation Based on Corrective Training and Learning Vector Quantization

Myoung-Wan Koo (1), Chong-Kwan Un (2)

(1) Basic Research Section 1, Korea Telecom Research Center, Seoul, Korea
(2) Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Korea

In this paper, we present a speaker adaptation technique based on corrective training(CT) and learning vector quantization (LVQ). Our algorithm consists of two stages: codebook adaptation and hidden Markov model(HMM) parameter adaptation. In the stage of codebook adaptation, we propose a codebook adaptation scheme using a neurally-inspired LVQ with highly discriminant ability. In the stage of HMM parameter adaptation, we propose a modified corrective training algorithm for speaker adaptation in which the HMM parameter adaptation obtained by probability transformation matrix arc re-estimated to maximize the recognition rate on the adaptation speech. With this method, the recognition rate for new speakers can be improved.

Full Paper

Bibliographic reference.  Koo, Myoung-Wan / Un, Chong-Kwan (1992): "A speaker adaptation based on corrective training and learning vector quantization", In ICSLP-1992, 1475-1478.