Sixth European Conference on Speech Communication and Technology
A recognition strategy that can be matched to specific system performance criteria such as word error rate or F-measure has recently been found to yield improvements over the usual maximum aposteriori probability strategy   . In this matched-to-the-task strategy a hypothesis is chosen to minimize the expected loss or the Bayes Risk under a loss function defined by a performance measure of interest. Due to the prohibitive of exact implementation of this strategy, only an approximate implementation as an N-best list rescoring scheme been used  . Our goal is to improve the performance of such risk-based decoders by developing search strategies that can consider more hypotheses and incorporate more acoustic evidence. In this paper we present search algorithms to implement the risk-based recognition strategy over word lattices that contain acoustic and language model scores. These algorithms are extensions of the N-best list rescoring approximation and are formulated as A* algorithms. Results are reported on the Switch-board conversational telephone speech corpus. We find that lattice based rescoring yields modest but significant improvements in word error rate relative to N-best list rescoring at comparable computational cost.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Goel, Vaibhava / Byrne, William (1999): "Task dependent loss functions in speech recognition: a* search over recognition lattices", In EUROSPEECH'99, 1243-1246.