EUROSPEECH 2001 Scandinavia
This paper proposes a new fast match algorithm for large vocabulary continuous speech recogniton. By viewing the fast match as a verification problem we develop a likelihood ratio score to be used instead of conventional likelihood based fast matches. We also improve the computation through incremental calculation and the use of SSE for Intel instruction set. When used in a 20K Japanese broadcast news task the proposed fast match leads to about 30-40% improvement in speed with almost no degradation in the recognition accuracy.
Bibliographic reference. Liu, Feng / Afify, Mohamed / Jiang, Hui / Siohan, Olivier (2001): "A new verification-based fast match approach to large vocabulary speech recognition", In EUROSPEECH-2001, 851-854.