Coloring the speech utterance to accelerate the SM based LVCSR decoding

This paper proposes a coloring speech utterance algorithm to speed up the segment based decoding process. It divides the acoustic space into several parts instead of separating the utterance into pieces, so the successive speech decoder can focus on some of these sub-spaces and the efficiency of the recognition system would be improved. Since the coloring speech algorithm is frame based, the results can be used by the decoder immediately. The coloring speech algorithm is successfully applied to a segment model based mandarin LVCSR system. Experiments show the run-time of the decoder is decreased by 2 times without obvious influence to the accuracy.