Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Predictive Speaker Adaptation Based on Least Squares Method

Woo-Yong Choi (1), Hyung Soon Kim (2)

(1) SLP-1 Team, SLP Division, L&H Korea
(2) Dept. of Electronics Engineering, Pusan National University, Korea

In this paper, we considered two representative speaker adaptation (SA) approaches - MAP and MLLR techniques for our Korean isolated word recognition task. In addition, we proposed a new speaker adaptation algorithm to improve the performance of MAP technique. It is based on least squares method between MAP adapted mean vectors and the corresponding SI mean vectors. The results of our experiment using a CDHMM system indicated that the proposed SA technique yielded high performance improvement in the recognition rate espe-cially when the number of adaptation data is very limited. Moreover, the computational load of proposed technique is much smaller than that of MLLR.


Full Paper

Bibliographic reference.  Choi, Woo-Yong / Kim, Hyung Soon (2000): "Predictive speaker adaptation based on least squares method", In ICSLP-2000, vol.3, 865-868.