This paper describes a real-time large-vocabulary telephone dialogue system. The system is designed for providing Voice-Activated Telephone Extension (VATEX) services for a company of about. 5000 employees. The system recognizes a user's request sentences and, if necessary, asks questions to specify one person then connects the line to the called person's branch telephone.
To realize fast and accurate continuous telephone speech recognition for large vocabulary applications, we developed new methods. To reduce errors of speech endpoint detection, we adopt a new speech endpoint detection algorithm which does not use speech energy level but likelihood of partial matching paths for the system. We also adopt- a two-level search algorithm and semantical merging method for the system.
The effectiveness of these algorithms was evaluated by recognition experiments using 300 sentences recorded over the telephone network. Using the speech endpoint detection method, degradation of recognition accuracy due to failure of endpoint detection is very small even at the SNR of 7 dB where speech detection using speech level does not work at all. The two-level search and semantical merging method reduces errors by 30 % and process time increases to about 2.5 % of the length of the user's utterance. Totally, the system can achieve 91 % task accuracy for telephone input.
Cite as: Naito, M., Kuroiwa, S., Takeda, K., Yamamoto, S., Yato, F. (1995) A real-time speech dialogue system for a voice activated telephone extension service. Proc. ESCA Workshop on Spoken Dialogue Systems, 129-132
@inproceedings{naito95_sds,
author={Masaki Naito and Shingo Kuroiwa and Kazuya Takeda and Seiichi Yamamoto and Fumihiro Yato},
title={{A real-time speech dialogue system for a voice activated telephone extension service}},
year=1995,
booktitle={Proc. ESCA Workshop on Spoken Dialogue Systems},
pages={129--132}
}