Confidence estimation of the output hypothesis of a speech recognizer offers a way to assess the probability that the recognized words are correct. This work investigates the application of confidence scores for selection of speech segments in unsupervised speaker adaptation. Our approach is motivated by initial experiments that show that the use of mis-labeled data has a significant cost in the performance of particular adaptation schemes. We focus on a rapid self-adaptation scenario that uses only a few seconds of adaptation data. The adaptation algorithm is based on an extension to the MLLR transformation method that can be applied to the observation vectors. We present experimental results of this work on the ARPA WSJ large vocabulary dictation task.
Cite as: Anastasakos, T., Balakrishnan, S.V. (1998) The use of confidence measures in unsupervised adaptation of speech recognizers. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0599, doi: 10.21437/ICSLP.1998-746
@inproceedings{anastasakos98_icslp,
author={Tasos Anastasakos and Sreeram V. Balakrishnan},
title={{The use of confidence measures in unsupervised adaptation of speech recognizers}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0599},
doi={10.21437/ICSLP.1998-746}
}