Study And Hardware Implementation Of Speech Enhancement Algorithm Based On Modulation Parameters In Recurrent Neural Network

With the development of wearable devices,video conferencing communication,humancomputer interaction,and other fields,speech enhancement technology has attracted extensive attention from scholars.Traditional speech enhancement algorithms in the time-frequency domain have the problems of a single noise estimation model,poor anti-interference ability,and poor speech enhancement effect.Although the speech enhancement method based on deep neural network learning can solve the above problems,its application in the embedded platform is limited due to the complexity of model weight parameters.To solve these problems,this paper first discusses the energy relationship between the pure speech signal and the same frequency noise signal in the training data.A new frequency band gain parameter extraction algorithm is proposed to optimize and improve the existing speech enhancement algorithm based on neural network learning,and then an application test platform is built based on FPGA to evaluate the performance of the designed speech enhancement algorithm.In the process of feature parameter extraction,the energy spectrum depth modulation algorithm is used to extract the frequency band gain parameters of the training set to solve the problem of over-estimation or under-estimation.Given the problem of complex algorithm models and slow calculation speed,this paper adopts the data processing method of frequency band gain estimation interpolation and the training method of shallow GRU network structure to improve the running speed of the algorithm in this paper.To verify the effectiveness of the proposed algorithm,we set up an experiment to test the speech enhancement performance of the proposed algorithm and existing algorithms from other papers.Compared with some existing algorithms in different noise environments,the results show that in the SNR(signal to noise ratio)range of 0d B-15 d B,under the interference of Babble,the average score of speech perception quality evaluation(PESQ)of this method is increased by 7.95%;Under the influence of Gaussian white noise,the average PESQ score of the proposed algorithm increases by 9.77%;Under Factory noise,the average PESQ score of the proposed algorithm increases by 7.19%;Under the influence of Pink noise,the average PESQ score of the proposed algorithm increases by 10.19%.Finally,an algorithm testing platform is built based on FPGA to test the hardware resource cost of the algorithm.The result shows that the total logical resource is occupied by 13.82%.