A Speech Recognition System Based On Statistical Model Research And The Realization Of Dsp

Speech recognition is used to identify the semantics of natural human voice, orused to identify who the speaker is by the characteristics of human voices. With thedevelopment of speech recognition systems, speech recognition technology has beenwidely applied to many fields such as medical, military, aerospace, mobile internet. Inrecent years, with the key technology to make breakthroughs, embedded speechrecognition system is applied to many consumer electronic products, it has profoundlychanged the traditional human-computer interaction model. Recognition accuracy androbustness is the key of the Speech Recognition System. In this thesis, from these twoperspectives, we will research the basic algorithm of the isolated word speechrecognition system and the OOV rejection algorithm for rejection processing and theimplementation of the system on the DSP platform.Firstly, this thesis discussed detailed description of the basic principles of speechrecognition systems and implementation technique, we focused on the front-endprocessing of speech signals, including the endpoint detection, extraction of speechcharacteristic parameters. Then we discussed the establishment and realization of HMM,mainly focused on the initialization of the HMM and how to merge the templateparameter.Secondly, the recognition result of speech recognition systems is always difficult toavoid the false consciousness that this will seriously affect the robustness of the systemand the recognition accuracy, So we need a way to reject OOV word. taking intoaccount the complexity and cost of the system implementation, we research theback-end processing technology of the speech recognition system, the principles andrealization of word rejection. We choose a rejection algorithm based on a posterioriprobability and LVQ, and Then we propose several characteristic parameters which canbetter differ OOV from IV by the posterior probability for word rejection. Merging theclass label and feature parameters as the input vector, and put the vector into LVQnetwork for training the network. Then the LVQ network has the ability to distinguishbetween the OOV and IV. Finally, training the network with different input vectors, as well as using the different test set to test the system rejection capability, and gives theIV rejection rate and OOV acceptance rate under different circumstances. The resultshows that the system would rejected about2.6%of the IV voice, at the same time, itcan reject the OOV voice more than98%.After the implementation of the system-related algorithm on the PC platform, westudy how to implement an isolated word speech recognition system on the DSPplatform. First we discuss the processor architecture and memory configuration on theDSP platform. In order to establish an efficient hardware platform, we study the internalconnections between the various chips, as well as various interface settings, the use ofaudio processing chip and the progress of collecting audio data which are particularlydiscussed in detail. Also we describe the software design process, and how to transplantthe speech recognition algorithm to the DSP platform from the PC platform. The systemmust be able to run without the emulator and development environment, so we discussthe boot-loader of the system. Finally, we get the general isolated word speechrecognition system based on DSP.