【摘要】 在大词表孤立词语音识别中,Viterbi搜索是时间消耗的主要因素。为改善基线系统性能,根据汉语孤立词识别的特点,提出了一种基于音节切分的束搜索算法,在音节层和词条层进行剪枝。该算法不增加内存开销。实验结果表明:在词表规模为10 000时,该算法以0.23%的识别率下降率为代价,将Viterbi搜索的时间消耗降低为基线系统的26.73%;相对于小词表,该算法在大词表情况下对系统性能的改善尤为明显。更多还原

【Abstract】 In large vocabulary isolated word speech recognition,the Viterbi search is the major time consumes.The performance of the baseline system is improved using the characteristics of isolated Mandarin words in a syllable detection based beam search algorithm with pruning of syllable and word levels.This algorithm has no additional memory cost.Tests show that for vocabulary with 10 000 words,the algorithm reduces the time consumption to 26.73% of the baseline system,while the recognition rate is reduced by only 0.23%.The performance is improved even more when the vocabulary contains more words.更多还原