Realization Of Robust Speech Recognition Algorithm Based On Two-layer GMM Structure And Multi-environment Optimization Model On ARM Platform

The embedded system has the advantages of low cost,small size and low energy consumption.Therefore,the speech recognition system based on the embedded hardware platform has a wide application.Vector Taylor Series(VTS)is a commonly used feature compensation for speech recognition,which is robust to noise and can effectively improve the recognition performance in the severe environment.However,since VTS feature compensation has iterative structure and exponent computation,the computation complexity of VTS is high.Therefore,this thesis first optimizes the structure of VTS feature compensation,including the use of two-layer GMM(Gaussian Mixture Model)model instead of the original single-layer GMM,and optimizes the parameter setting based on the ARM A8 platform to.Then,multi-model VTS algorithm,based on ARM A8 platform is also proposed to match optimal environment,which improves the adaptability of recognition system to complex acoustic environment.The main work of this thesis is summarized as follows:(1)An overview of speech recognition system based on ARM platform.This thesis analyzes the basic process and structure of speech recognition system,including preprocessing,feature extraction and HMM(Hidden Markov Model)matching.(2)Realization and optimization of the two-layer GMM VTS based speech recognition system on ARM platform.In single-layer GMM based VTS feature compensation algorithm VTS feature compensation module occupy the majority of computation of the speech recognition system.VTS feature compensation with single-layer GMM is analyzed on MATLAB platform and ARM platform respectively.According to the simulation test results,the optimal mixing number of the each GMM is determined.Also,this computation structure of VTS is optimized.Based on the simulation results of ARM platform,it is shown that the computation of two-layer GMM structure is greatly reduced,and the recognition performance is also comparable to that of the original single-layer GMM.(3)Realization and optimization of two-layer GMM structure and multi-noise environment model VTS speech recognition system on ARM platform.The original multi-environment model based VTS speech recognition has the better performance in white noise environment,while has deteriorated performance for other noises environment.Thus,this thesis introduces the training model under several types of noise,and selects the SNR categories needed for the training model.The simulation experiment based on MATLAB and ARM platform show that the optimized multi-noise environment system improves the recognition rate of the system in a variety of noise environments.At the same time,the tests on the ARM platform show that the optimization method of this thesis not only reduces the storage capacity of the system,but also reduces the time it takes for the best model matching part.