In this paper we introduce an effective speaker adaptation technique to our unique compact speech recognizer especially designed for consumer electronics products. The compact ASR method we have developed in our previous work employs phoneme similarities as feature parameters, which are extracted temporally successive matching between speech sample and 24 context-independent phoneme standard templates in forms of time-spectral pattern. With these unique feature parameters, the recognizer has been successfully embedded in a single DSP without any external memory chips. Speaker adaptation method we propose in this paper focuses on adaptation of the phoneme standard templates in the ASR method. When evaluated with 20 or less training words in 100 isolated words recognition test, the proposed method achieved reduction of error rate by 20 % approximately. The test results also show its effect for child subjects and validity of applying this technology in the recognizer which is used by various users in consumer electronics products.
Cite as: Ono, Y., Yamada, M., Hoshimi, M. (1999) A study of speaker adaptation for speaker independent speech recognition method using phoneme similarity vector. Proc. 6th European Conference on Speech Communication and Technology (Eurospeech 1999), 2523-2526, doi: 10.21437/Eurospeech.1999-552
@inproceedings{ono99_eurospeech,
author={Yoshio Ono and Maki Yamada and Masakatsu Hoshimi},
title={{A study of speaker adaptation for speaker independent speech recognition method using phoneme similarity vector}},
year=1999,
booktitle={Proc. 6th European Conference on Speech Communication and Technology (Eurospeech 1999)},
pages={2523--2526},
doi={10.21437/Eurospeech.1999-552}
}