| Due to the development of chip technology and big data,artificial intelligence technology has been fully developing.Therefore,the application of voice command word recognition technology based on deep learning is becoming more and more widespread.However,in practical applications,whether based on traditional technology or deep learning technology,the command word recognition still has the problem of insufficient robustness in the practical application environment,that the recognition accuracy will be severely reduced if disturbed by noise in the practical application.In this paper,the traditional noise reduction algorithm of microphone array and neural network are used as the main technologies,and some methods to improve the robustness and accuracy of speech command word recognition are studied and proposed.An optimized logarithmic minimum mean square error speech enhancement algorithm based on equilateral triangle array is proposed for the interference of directional noise and speech distortion caused by classical logarithmic minimum mean square error algorithm that reduces the back-end speech recognition system performance.Selecting any microphone data in the array as the reference,performing subtraction operation with the data of the other two microphones in turn and reserving the results to get the mean value,can effectively suppress the interference signals outside the front of the reference microphone.At the same time,the logarithmic minimum mean square error algorithm is used as the post-filtering algorithm that gain function and noise estimation method are optimized.Finally,the speech high frequency signal distortion after noise reduction processing is compensated by proposed high frequency compensation algorithm.Experimental results indicate that compared with the classical logarithmic minimum mean square error algorithm,the SNR range for improving of the proposed algorithm is0.13 d B～5.08 d B under different SNR conditions in multiple environments.The score range of improving PESQ is 0～0.28.A multichannel command word recognition system based on power normalized cepstrum coefficient and neural network is proposed to solve the problem of insufficient robustness of command word recognition under non-user speech and noise interference.Firstly,multi-channel array is used to record the voice command word datasets with rich spatial features.Secondly,the power normalized cepstrum coefficient,which is robust to noise,is used as the input of the command word recognition model.Finally,three multitask command word recognition models are constructed by residual shrinkage network and inception residual network modules respectively,which can recognize the command word and determine whether the word is issued by the user.In the command word recognition task,the accuracy rates of the proposed Res Nets-CW and Multi-IRes Nets models are93.39% and 92.97% respectively under the condition of adding noise.In addition,the accuracy of the constructed low-power model LRes Nets-CW is 91.25%.When the model parameters are greatly reduced,it can still maintain high performance,which greatly reduces the power consumption of the system and is more suitable for deployment on lowpower intelligent devices.At the same time,the accuracy of the user discrimination function of the three models in the test data set is over 95%.The system modules of the front-end speech enhancement algorithm proposed in this paper are designed and further implemented by software and hardware platforms,and its performance in practical applications is evaluated by testing in real environment,and the effectiveness of the algorithm is also verified.It is worth noting that through cooperation with a medical device company focusing on hearing research in Shanghai,the algorithm has been successfully applied in the field of hearing rehabilitation of medical device,which deployed in front-end auxiliary hearing device coordinate work with cochlear implants and hearing AIDS to provide better listening experience for the hearing-impaired group. |
PDF Full Text Request