小资源下语音识别算法设计与优化  

Design and optimization of a low resource speech recognition system

作　　者：张鹏远[1] 计哲[2] 侯炜[2] 金鑫[2] 韩卫生[1] ZHANG Pengyuan JI Zhe HOU Wei JIN Xin HAN Weisheng(Key Laboratory of Speech Acoustics and Content Understanding, Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China National Computer Network Emergency Response Technical Team/Coordination Center of China, Beijing 100029, China)

机构地区：[1]中国科学院声学研究所,语言声学与内容理解重点实验室,北京100190 [2]国家计算机网络应急技术处理协调中心,北京100029

出　　处：《清华大学学报（自然科学版）》2017年第2期147-152,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金项目(U1536117,11590770,11590771,11590772,11590773,11590774);中国科学院战略性先导科技专项(XDA06030100,XDA06030500,XDA06040603);国家“八六三”高技术项目(2015AA016306);国家“九七三”重点基础研究发展计划项目(2013CB329302)

摘　　要：可穿戴设备和智能家居系统需要语音识别引擎占用极小的资源并具有较强的拒识能力。传统的语音识别算法无法满足小资源系统的这种需求。该文针对小资源下语音识别系统,在解码策略和拒识算法设计上均提出了改进方法。在解码策略上,通过修改垃圾音素的重入,使得集外语音的拒识率提高到64.8%,而内存占用只增加了8.5kB。在拒识算法上,提出了离线计算背景概率和在线查表的方法,与基线系统相比,在集内识别率略有损失的情况下,集外拒识率达到93.8%,而内存占用和计算速度也得到了优化。Wearable devices and smart home systems need speech recognition engines with few resources and high rejection rates. Traditional methods cannot provide such systems, This paper presents algorithms for decoding and rejection for a low source speech recognition system. The decoding improves the rejection rate up to 64.8% by changing the filler reentry while the memory is only increased 8.5 kB compared with the baseline system. The rejection algorithm computes a background probability which is compared to similar probabilities calculated in advance online decoding. The system gives a rejection rate of 93.8- with little loss in the recognition rate. The memory and computational speed are also optimized.

关 键 词：语音识别 小资源 置信度 

分 类 号：TN912.34[电子电信—通信与信息系统]