基于多语音增强与分离的对话机器人系统设计  

Research on application of short video recommendation model based on collaborative filtering algorithm

作　　者：袁芝妹 张华 王丽 YUAN Zhimei;ZHANG Hua;WANG Li(Hunan College of Foreign Studies,Changsha 410000,China)

机构地区：[1]湖南外国语职业学院,长沙410000

出　　处：《自动化与仪器仪表》2024年第10期282-286,共5页Automation & Instrumentation

基　　金：2021年度湖南省社会科学成果评审委员会课题《人工智能背景下高职外语类专业人才培养转型研究》(XSP21YBC363)。

摘　　要：为进一步提升英语语音交互机器人在复杂环境下的语音交互效果,基于语音分离技术,进行英语语音增强方法的设计,以进一步提升人机英语语音交互的效果。其中,以DPRNN作为基础的语音分离模型,并以此为基础进行优化,最终通过自适应注意力进行最终的优化,以进一步提升语音分离的效果。实验结果表明,与Conv-TasNet和DPRNN相比,改进后的DPRNN具有更快的收敛速度,同时还能保持损失较低;与其他分离模型相比,设计的TAANet语音分离方法具有更好的语音分离效果,更强的适应能力与泛化能力,在WSJ0、WHAM、WHAMR以及GRID-2mix英语语音数据集上的SI-SNRi和SDRi得分分别达到了25.94和24.30、15.55和15.81、13.03和11.23、16.49和17.12;同时,设计的TAANet英语语音分离方法在各种噪声环境下的PESQ得分均能得到不同程度的提升,稳定性较强。以上结果表明,设计的TAANet英语语音分离方法能够实现效果良好的语音分离,将其应用于人机交互的语音增强是可行的,可靠性较高。In order to further improve the speech interaction effect of English speech interaction robots in complex environments,based on speech separation technology,an English speech enhancement method is designed to further enhance the effectiveness of human-machine English speech interaction.Among them,a speech separation model based on DPRNN is optimized based on it,and finally optimized through adaptive attention to further improve the effectiveness of speech separation.The experimental results show that compared with Conv TasNet and DPRNN,the improved DPRNN has faster convergence speed while maintaining lower losses;Compared with other separation models,the designed TAANet speech separation method has better speech separation performance,stronger adaptability and generalization ability,and achieves SI-SNRi and SDRi scores of 25.94 and 24.30,15.55 and 15.81,13.03 and 11.23,16.49 and 17.12 on WSJ0,WHAM,WHAMR,and GRID-2mix English speech datasets,respectively;At the same time,the designed TAANet English speech separation method can achieve varying degrees of improvement in PESQ scores in various noise environments,with strong stability.The above results indicate that the designed TAANet English speech separation method can achieve good speech separation,and its application in speech enhancement in human-computer interaction is feasible and reliable.

关 键 词：语音交互 语音增强 语音分离 DPRNN 自适应注意力网络 

分 类 号：TP392[自动化与计算机技术—计算机应用技术]