Sixth International Conference on Spoken Language Processing
This paper describes an enhanced pruning technique aimed at a further reduction of the active search space in large vocabulary speech recognition, to speed-up decoding while maintaining the accuracy. The method is based on anticipating both the linguistic and acoustic contribution of a phonetic arc, before expanding that arc in the search. The decoder is based on a time-synchronous beam search and a lexical tree. Cross-word HMMs and M-gram language models are integrated in a single decoding pass. The new algorithm has been evaluated for one-pass trigram decoding of Broadcast news. With respect to the baseline, the search effort can be halved at almost no degradation. When pruning more aggressively to get a speed-up of 10, real-time decoding is achieved on Hub4 evaluation, however, with an increase of the base error rate by one third.
Bibliographic reference. Aubert, Xavier L. / Blasig, Reinhard (2000): "Combined acoustic and linguistic look-ahead for one-pass time-synchronous decoding", In ICSLP-2000, vol.3, 802-805.