Embedded Speech Recognition System Based On Dsp And Dhmm Research And Implementation

In recent years, with the development of speech recognition technology, speechrecognition system based on embedded systems is widely applied to the field of smarttoys, industrial control, and medical services, which leads to convenient services forpeople’s lives. To realizing efficient speech recognition system on the embeddedplatform has become a hot topic owing to that embedded systems advantages in cost,size, and power consumption. In the field of isolated word speech recognition systembased on the DSP embedded system platform, there needs to do deeper research todevelop some simple and applicable speech recognition system.This thesis studies the system design of embedded speech recognition systemwhich based on the DSP and DHMM, and ultimately implements a non-specific personisolated word speech recognition system. The main work includes:First, the theory and the key technologies of the speech recognition system areresearched. From the pre-processing analysis of speech signals, this thesis graduallyanalyses the theory and implementation of endpoint detection, the phonetic featuresparameters extraction and vector quantization; and summarizes the advantages anddisadvantages of typical template matching algorithm in speech recognition.Second, the basic principles and three basic questions of hidden Markov model aredescribed. By analyzing the characteristics of the hidden Markov model, the discretehidden Markov model is selected as the matching algorithm for isolated word speechrecognition. This thesis studies the calibration in model training, multiple observationsequences parameters revaluation algorithm, as well as the observation symbolprobability distribution matrix processing in training of the discrete hidden Markovmodel. At the same time, this thesis solves the problem that individual isolated wordrecognition rate is not high through defining the learning function of the isolated words.Third, the hardware platform for embedded speech recognition system isconstructed by taking the DSP processor TMS320VC5509A as the core hardwarecomponents. Through researching and analyzing the hardware platform resourcelimitations and the characteristics of the speech recognition system based on DHMM, the system software architecture consist of the embedded system software andPC-assisted software is designed; and selects USB as the data communication tool ofthe two software systems. The main function of the embedded system software iscompleting the speech recognition function, while the additional function is doingsimple training and learning. Moreover, the functions of best codebook training, voicetemplates training and learning, and data updating are completed on PC-assistedsoftware. This thesis describes the design and implementation each module of thesoftware.Finally, this thesis studies the optimization of embedded system software. Thereal-time response speed of the system is reduced from about10seconds to an averageof230milliseconds by optimizing the software algorithm in fixed-point and designingthe system storage in efficient way. Through the optimization of the Viterbi algorithm,real-time response speed of the system once again increased by36.4%and can finishrecognition in157.4milliseconds in average. At the same time, the recognition rate ofcan reach90%with the vocabulary within100.