计算机辅助普通话测试系统中基于神经网络的语音自动增强方法研究  

Research on Automatic Speech Enhancement Method Based on Neural Network in Computer Aided Putonghua Testing System

作　　者：唐文静[1] TANG Wenjing(Shangluo Vocational and Technical College,Shangluo Shanxi 726000,China)

机构地区：[1]商洛职业技术学院,陕西商洛726000

出　　处：《自动化与仪器仪表》2023年第5期70-73,共4页Automation & Instrumentation

基　　金：商洛职业技术学院院级重大课题《地方高职院校语言文字规范化建设的路径方法与实践研究—以商洛职业技术学院》(JYKT2022004)。

摘　　要：为提升普通话测试的效果,提出一种基于神经网络的语音自动增强方法,以提高测试语音的清晰度。通过将独立自注意力机制与全卷积神经网络进行融合,以解决全卷积神经网络存在的感受域内放缩特性差的问题,提升语音增强方法的整体性能。结果表明,基于独立自注意机制融合全卷积神经网络的语音增强算法SAUNet能够取得良好的语音增强效果,与AAUNet和UNet两种语音增强方法相比,在评价指标SIG、BAK、OVL、PESQ以及STOI上分别提高了17.4%和8.7%、13.44%和8.27%、15.47%和1054%、15.45%和7.58%、9.25%和6.88%。实验结果表明本研究提出的语音增强方法的性能优越,能够使得计算机辅助下的普通话测试取得更好的测试效果,具有一定实际参考设计价值。In order to improve the comprehensive effect of the current computer-assisted Putonghua test,an automatic speech enhancement method based on neural network is constructed.By fusing the independent self attention mechanism with the full convolutional neural network,the shortcomings of the full convolutional neural network can be solved,and the overall performance of the speech enhancement method can be improved.The simulation results show that SAUNet,a speech enhancement algorithm based on independent self attention mechanism proposed in this study,can achieve good speech enhancement effect.Compared with AAUNet and UNet,SAUNet has improved 17.4% and 8.7%,13.44% and 8.27%,15.47% and 1054%,15.45% and 7.58%,9.25% and 6.88% respectively in the evaluation indicators SIG,BAK,OVL,PESQ and STOI.The experimental results show that the speech enhancement method proposed in this study has superior performance,which can make the computer assisted Putonghua test achieve better test results,and has certain practical reference design value.

关 键 词：普通话测试 卷积神经网络 语音增强 注意力机制 

分 类 号：TP392[自动化与计算机技术—计算机应用技术]