 |
ITRW on
|
 |
|
ITRW on
|
|
In this paper, we propose a new approach to hidden Markov model (HMM) adaptation based on the probabilistic principle component analysis (PPCA). The proposed approach has been developed to adapt not only the HMM means but also the variances and mixture weights simultaneously. Due to a set of constraints, we apply the PPCA model in the transformed domain where we adapt the variance and mixture weight. The PPCA model provides the information of correlation among speech units as well as the prior probability density function (pdf) associated with each HMM parameter. In order to adapt the HMM parameters, we use the generalized expectation maximization (GEM) algorithm in which the M-step requires the parameters to be chosen such that the auxiliary function could be increased. The experimental results show that the proposed PPCA-based adaptation approach significantly outperforms the conventional MAP approach when only a small amount of adaptation data is provided.
Bibliographic reference. Kim, Dong Kook / Kim, Nam Soo (2001): "Maximum a posteriori adaptation of HMM parameters based on probabilistic principle component analysis", In Adaptation-2001, 25-28.