Tree-trellis forward-backward algorithm has been widely used for N-best search in continuous speech recognition. In conventional approaches, the heuristic score used for the A* backward search is derived from the partial-path scores recorded during the forward pass. The inherently delayed use of language model in the lexical tree structure leads to inefficient pruning and the partial-path scores recorded is an underestimated heuristic score. This paper presents a novel method of computing the heuristic score that is more accurate than the partial-path score. The goal is to recover high-score sentence hypotheses that may have been pruned halfway during the forward search due to the delayed use of LM. For the application of Hong Kong stock information inquiry, the proposed technique shows a noticeable performance improvement. In particular, a relative error-rate reduction of 12% has been achieved for top-1 sentences.
[1]
R. Schwartz,et al.
A comparison of several approximate algorithms for finding multiple (N-best) sentence hypotheses
,
1991,
[Proceedings] ICASSP 91: 1991 International Conference on Acoustics, Speech, and Signal Processing.
[2]
Frank K. Soong,et al.
The use of tree-trellis search for large-vocabulary Mandarin polysyllabic word speech recognition
,
1994,
Comput. Speech Lang..
[3]
Frank K. Soong,et al.
A Tree.Trellis Based Fast Search for Finding the N Best Sentence Hypotheses in Continuous Speech Recognition
,
1990,
HLT.
[4]
Tan Lee,et al.
Lexical tree decoding with a class-based language model for Chinese speech recognition
,
2000,
Interspeech.