In this paper, we examine the use of speaker adaptation in adverse noise conditions. In particular, we focus on incremental on-line speaker adaptation since it, in addition to its other advantages, enables joint speaker and environment adaptation. First, we show that on-line adaptation is superior to off-line adaptation when realistic changing noise conditions are considered. Next, we show that a conventional left-to-right HMM structure is not well suited for on-line adaptation in variable noise conditions due to unreliable state-frame alignments of noisy utterances. To overcome this problem, we suggest the use of state duration constrained HMMs. Our experimental results indicate that the performance gain due to adaptation is much greater with duration constrained HMMs than obtained with conventional left-to-right HMMs. In addition to the appropriate model structure, we point out that in long-term adaptation, such as incremental on-line adaptation, the supervised approach is a necessity.
Cite as: Viikki, O., Laurila, K. (1998) Incremental on-line speaker adaptation in adverse conditions. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0313, doi: 10.21437/ICSLP.1998-742
@inproceedings{viikki98_icslp,
author={Olli Viikki and Kari Laurila},
title={{Incremental on-line speaker adaptation in adverse conditions}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0313},
doi={10.21437/ICSLP.1998-742}
}