基于连续识别的嵌入式孤立词识别系统  

An Isolated-word Speech Recognition System Based on CSR for Hand-Held Devices

作　　者：冷冰涛[1] 梁维谦[2] 董保帅[1] 原道德[2] 

机构地区：[1]清华大学微电子学研究所,北京100084 [2]清华大学电子工程系,北京100084

出　　处：《电声技术》2011年第11期42-45,共4页Audio Engineering

基　　金：国家863高技术项目(863计划)重点项目(2008AA010700)

摘　　要：基于线性网络的孤立词识别系统识别时间与词表规模成正比,识别性能严重受限于词表的规模。根据汉语孤立词特点,提出了一种基于连续识别的大词表孤立词识别系统。该系统围绕嵌入式识别所关注的速度和内存消耗性能,采用先实现应用多级搜索、定点化策略的连续识别,再对连续识别的识别结果进行音字转换处理的方法,将连续语音识别应用于大词表孤立词识别任务。在10 000词表规模下该系统与采用线性网络的基线系统相比,识别率基本保持不变,识别时间和内存消耗下降到10%左右。Traditional isolated words recognition＇ s decoding speed depended on the size of vocabulary, which is based on linear net. There would be a serious increase on decoding time with vocabulary＇ s increasment. According to Chinese characteristic, an Mandarin isolated words recognition system based on continuous speech recognition is proposed in this paper. At continuous recognition step, it uses multi - pass decoding and fixed point computation strategy, and then it handles the syllable string to get Mandarin isolated words. With 10 000 words there is little decrease in accuracy and 90 percent improvement in recognition speed and memory consumption.

关 键 词：大词表孤立词识别 嵌入式 连续语音识别 音字转换 

分 类 号：TN912[电子电信—通信与信息系统]