基于HTK的语音识别的并行化研究与实现  

Research and implementation of parallel speech recognition based on HTK

作　　者：刘勇进[1] 史晓东[2] 

机构地区：[1]厦门大学计算机科学系,福建厦门361005 [2]厦门大学智能科学与技术系,福建厦门361005

出　　处：《计算机应用》2009年第4期1052-1055,共4页journal of Computer Applications

基　　金：国家自然科学基金资助项目(60573189);国家863计划项目(2006AA01Z1392006AA010108-3);福建省重点科技项目(2006H0038);福建省自然科学基金项目(2006J0043)

摘　　要：详细地分析了语音识别的过程,给出了相应的算法描述,并分析了语音识别并行化的可能性。将并行计算的思想应用于语音识别的算法中,使用多线程技术,并引入避免竞争条件的机制,在多核计算机上并行地计算HMM模型节点的似然率,从而得到语音识别的并行化算法。分析了该并行化算法的性能,同时在语音识别工具包HTK 3.4上实现了这种并行化算法。基于WSJ0语料库的实验结果表明该并行化算法在不影响识别结果的前提下能够有效地提高语音识别的实时性能。After comprehensively analyzing the process of speech recognition and depicting its corresponding algorithm, the feasibility of the parallelism in the algorithm was analyzed. The parallel computing concept and the multi-threading technology were applied in the algorithm of speech recognition, and a protection mechanism was introduced to avoid the occurring race condition during the computing of likelihood of Hidden Markov Model （HMM） on muhi-core computer. Then the parallel algorithm for speech recognition was proposed and its performance was also evaluated. The experiments on WSJO corpora demonstrate that the implementation of parallel algorithm which was realized in HTK 3.4 toolkit can greatly improve the real-time performance of speech recognition without affecting the results of recognition.

关 键 词：多核 HTK 并行计算 语音识别 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]