Backward Viterbi Beam Search for Utilizing Dynamic Task Complexity Information

The backward Viterbi beam search has not received enough attentions other than being used in the second pass. The reason is that the speech recognition society has long ignored the concept of dynamic complexities of a speech recognition task which can help us to determine whether we should operate Viterbi decoding in forward or backward direction. We use the U.S. street address entry task as one example to show why the analysis of the dynamic task complexity information is important. We then describe the procedure to implement a backward Viterbi decoding system from an existing traditional ASR engine. We evaluated the backward Viterbi search using CMU's Pocket Sphinx. The experimental results of the backward Viterbi Search show significant performance improvement over the forward Viterbi decoding system on the U.S. street address entry task.