【作者】 刘雨辰； 徐明星；

【Author】 Yuchen Liu 1,Mingxing Xu 1 1.Key Laboratory of Pervasive Computing,Ministry of Education Tsinghua National Laboratory for Information Science and Technology(TNList) Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China

【机构】 普适计算教育部重点实验室清华信息科学与技术国家实验室(筹)清华计算机科学与技术系；

【摘要】 基于子词声学建模的关键词检出系统具有实时性好、词表可灵活配置的优点。在解码过程中,展开的词内搜索网络在词的首尾未能充分利用上下文相关的子词模型,而跨词的搜索网络具有难展开、展开规模大、复杂性高的缺点。本文在词内搜索网络上进行跨词搜索,通过反向搜索网络,逆向输入语音,与正向搜索交汇的双向搜索方法来降低剪枝风险。实验结果表明,跨词搜索算法相较于词内搜索有显著的性能提升,在两个测试集上分别相对提高27.8%、18.4%,双向搜索策略对有剪枝的跨词搜索有一定的性能提升,分别相对提高1.9%、1.7%。更多还原

【Abstract】 Higher real-time and the flexibility of word list are the advantages of the keyword spotting based on sub-word acoustic model.In decoding process,the use of context dependency sub-word acoustic model is not enough on the expanded word-internal search network,and the cross-word search network is hard to expand and will produce tremendous scale and large complexities after expanding.This article perform a cross-word decoding on the word-internal search-net,and try to decrease the beam-pruning risk by a bidirectional search,which search forward and backward simultaneously to meet in center.We do some experiments on a reading speech test set.The experiments show that the cross-word decoding algorithm has large advantages than word-internal decoder that get 27.8% and 18.4% relative improvement on two sets of test data respectively.The bidirectional search makes some improvements on the cross-word decoder with beam-pruning that get 1.9% and 1.7% relative improvement respectively.更多还原