语音识别开发工具包SRDK的研究与开发

Research and Development of Speech Recognition Development Kit

详细介绍了一个语音识别开发工具包SRDK(SpeechRecognitionDevelopmentkits)。该工具包可以方便地完成语音识别的各种任务,并且可以用来对语音识别技术进行研究。SRDK的特点是:ANSIC编写,便于向嵌入式系统进行移植;模块化良好,可以任意拆分组合;内置状态捆绑、训练中的剪枝、段长后处理、SSE(StreamingSingle-InstructionMultiple-DataExtensions)指令集的使用等多种先进技术等。已经使用SRDK开发出实用的语音识别系统。

Today,using the general-purpose platform to build speech recognizers becomes more and more popular.In this paper,a compact speech recognition development kit(SRDK)featured with effective merits is presented.SRDK is a set of software modules based on modified Hidden Markov Model(HMM).With them,the task for building various prac-tical speech recognizers as well as relative research work becomes more easily.SRDK is written in ANSI C,therefore it can run well not only on Windows operation system but also in UNIX environment.Further more it can be transplanted to other embed systems too.Contributed to the modularization design,any part of SRDK can be employed independently.Besides,in SRDK,four particular build-in approaches should be emphasized here.Firstly,parameter tying can be imple-mented in both semi-syllable level and state level.Secondly,SRDK adopts pruning in training stage in order to enhance the training speed.Thirdly,considering the effect of different speech rate,duration model is imported into the post pro-cessing for improving recognition performance.Finally,source codes of SRDK are optimized by using Streaming SIMD Extensions(SSE)instructions published by Intel Company and supported by AMD Company.Plus oriented graphs frame-works are utilized instead of multi-sub-tree structure in searching network,the recognition performance is improved comprehensively.The writers have already achieved a private automatic branch exchange system based on SRDK intro-duced in this paper.