Efficient decoding algorithms for Mandarin connected digit speech recognition

摘要

In this paper, a set of efficient decoding algorithms based on a speaker independent Mandarin connected digit speech recognition system is proposed. By simplifying the computation of observation probabilities and adopting an improved beam search pruning algorithm combined with duration information, the average decoding time is reduced from 0.92 second to 0.11 second per digit string while the recognition rate of unknown-length string is only decreased from 94.3% to 94.0%.