The Optimization And Implementation Of The Efficiency And Performance Of Chinese Language Model Based On Recurrent Neural Network

With the developing of artificial intelligence,smart home,intelligent medical,intelligent education will gradually enter our lives,work and study.In order to complete the interaction between human and machine,these artificial intelligences require a voice recognition system to support.The language model in the speech recognition system has a great impact on the efficiency and performance of the speech recognition system.Generally,it is hard to balance the efficiency and performance.Therefore,how to improve the efficiency and ensure that the performance does not decline or can be increased is in great significance.In this thesis,we mainly optimize the application of recognition accuracy and calculation speed in speech recognition process.This topic is based on “Speech recognition optimization based on RNN(Recurrent Neural Network)language model” project of the speech recognition group of IFLYTEK Co.,Ltd.In this thesis,the RNN language model is proposed by analyzing the shortcomings of N-Gram language model,but the high complexity of RNN language model limits its application in practical scene.After the intensive analysis and research,this thesis proposed a Fixed-point scheme based on RNN language model.The scheme used the SIMD(Single Instruction Multiple Data)instruction set to convert the float matrix operation of the RNN into the Fixed-point calculation,and combing with the project requirements,makes further optimization to improve the efficiency and performance of the model.In terms of minimizing the slightly fluctuations resulted from the improvement of performance,this thesis was put forward the pipeline performance optimization scheme after intensive study and analysis.By expanding the N-Best and merging the similar part of N-Best to form the prefix tree,then this thesis using the multi-threading technology to improve the efficiency and optimize the performance at the same time.In this thesis,the expanding N-Best results are rescored by RNN language model and N-Gram language model with interpolating the calculation.Then the results are reordered so that the optimal Top1 result can be choose as the final recognition result.Finally,the Fixed-point program and pipeline performance program were tested.The experimental results show that the proposed scheme improves the efficiency of RNN language model and ensures that there is no significant fluctuation.Then theexpanding N-Best were optimized by using multi-threaded prefix three to ensure that the improvement of the performance and the efficiency did not decline.The optimization scheme proposed in this thesis has been applied to the software of IFLYTEK Input Method Co.,Ltd.