一种高效的语音关键词检索系统  

Efficient keyword spotting system for information retrieval

作　　者：罗骏[1] 欧智坚[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《通信学报》2006年第2期113-118,共6页Journal on Communications

基　　金：国家自然科学基金资助项目(60402029)~~

摘　　要：针对音频检索任务中的关键词检索提出一种新的基于拼音图的两阶段检索系统,可以高效地从大量语音数据中检索出感兴趣的文本信息,从而达到为国家安全服务的目的。该系统分为预处理阶段和检索阶段。预处理阶段将语音数据识别成具有高覆盖率的拼音图,在这一过程中通过若干次的无监督最大似然线性回归自适应算法渐次提高拼音图的质量。检索阶段响应用户的频繁查询,只需在拼音图中查找出与关键词拼音匹配的拼音串,并采用基于N元拼音文法的前后向算法计算置信度以实现对检索结果的筛选。实验表明:系统具有较高的召回率和正确率,且检索阶段仅需0.01倍实时,可以满足快速检索的需要。A new two-stage keyword spotting system was proposed based on syllable graph for audio information retrieval task, which could efficiently spot the interesting words in mass speech data, thus serve for the national security. It comprised two stages - preprocessing and searching. In the preprocessing stage, the audio data was recognized into syllable graph which included high accuracy syllable candidates, and unsupervised MLLR （maximum likelihood linear regression） adaptation was carded out iteratively to further improve the accuracy of the syllable graph. In the searching stage, to answer the frequent queries from users, searching for matched keywords was only scanned in the graph for likely syllable strings. A forward-backward algorithm based on syllable N-grammar was used to calculate confidence measures for further filtering of the searching result. Experimental results show the system achieved good performances in both recall rate and accuracy rate, and in the searching stage only 0.01 times of real time is needed, which can meet the demand for fast retrieval.

关 键 词：信息检索 关键词检索 拼音图 置信度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]