基于语音识别的多模态人机交互系统关键技术研究

Research on key technologies of multimodal human-computer interaction system based on speech recognition

为了降低多模态人机交互系统在语音识别中的误差,研究提出了一种基于一维向量卷积神经网络(1-Dimensional Convolution Neural Network,1DCNN)的英语语音识别技术,并提出语音去噪方法,以此构建多模态人机交互系统。在1DCNN算法的性能分析中显示,研究采用的1DCNN算法损失值较小,且在训练中的误差值在0.005以内。此外在多模态人机交互系统的测试中显示,系统能够有效处理噪声信号,并且在不同情绪语音鉴别中的准确率较高,同时针对混合英语语音的鉴别准确率也在90%以上。以上结果表明,采用1DCNN算法能够有效提升多模态人机交互系统的语音识别和处理能力,为多模态人机交互的普及提供了设计方向,并为交互技术的未来发展提供理论支持。

In order to reduce the error of multimodal human-computer interaction system in speech recognition,an English speech recognition technology based on 1DCNN is proposed,and a speech denoising method is proposed to build multimodal human-computer interaction system.The performance analysis of 1DCNN algorithm shows that the loss value of 1DCNN algorithm used in the study is small,and the error value in training is within 0.005.In addition,the test of multimodal human-computer interaction system shows that the system can effectively process noise signals,and has a high accuracy rate in speech recognition of different emotions.At the same time,the accuracy rate for mixed English speech recognition is also above 90%.The above results show that the 1DCNN algorithm can effectively improve the speech recognition and processing ability of multimodal human-computer interaction system,and provide a design direction for the popularization of multimodal human-computer interaction,without providing theoretical support for the future development of interaction technology.