Sixth European Conference on Speech Communication and Technology
Multiple-pass search strategies are a necessary choice to reduce the computational cost and memory requirements of large vocabulary speech recognition systems uisng different kinds or complex modelling of speech and language. In the N-best approach the most efficient knowledge sources (usually acoustic models and a bi/trigram language model) are used first to select a short list (N-best) of alternative hypotheses. Then, the remaining more complex knowledge sources are used to rescore the sequences. In this paper we propose a new algorithm to overcome the problems associated with the simple ''Traceback-Based'' N-best algorithm while conserving its simplicity and low computational cost.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Fonollosa, A. R. / Batlle, Eloi (1999): "Combining length restrictions and n-best techniques in multiple-pass search strategies", In EUROSPEECH'99, 2713-2716.