基于深度学习的非稳态噪声抑制算法  

Transient Noise Suppression Algorithm Based on Deep Learning

作　　者：陈荣观 薛建清 陈东敏 CHEN Rongguan;XUE Jianqing;CHEN Dongmin(Fujian i-Star Technology Co.,Ltd.,Fuzhou 350002,China)

机构地区：[1]福建星网智慧科技有限公司,福建福州350002

出　　处：《电声技术》2020年第6期107-110,共4页Audio Engineering

摘　　要：在视频会议中,非稳态噪声(瞬时噪声)如敲击桌子的声音和鼠标敲击的声音,会分散听者注意力,干扰会议内容,影响视频会议质量。基于rnnoise神经网络模型,针对会议场景,采集办公场景非稳态噪声制作训练集,将输入48 kHz信号分为3个频带。其中:低频带采用深度学习模型降噪;中高频带使用低频带抑制系数加权平均,在确保算法效果的同时,进一步提升算法效率。在1.8 GHz ARM Cortex-A17 core的设备上,采用开源的音频库THCH-30、aidatatang_200h进行测试。结果表明:语音段中,增强效果12 dB及以上;非语音段中,增强效果24 dB及以上,同时保证语音没有明显不失真。对比rnnoise模型,每10 ms的处理时间缩短43.29%。In video conferences,transient noise,such as the sound of knocking on a desk or clicking the mouse,distracts the listener,interferes with the content of the meeting,and affects the quality of the video conference.In this paper,based on the rnnoise neural network model,for the conference scene,the common non-stationary noise is collected to make a training set,the input 48kHz signal is divided into three frequency bands,the low frequency band uses deep learning model to reduce noise,and the middle and high frequency bands use low frequency band suppression coefficient weighting on average,while ensuring the effect of the algorithm,it further improves the efficiency of the algorithm.On the 1.8GHz ARM Cortex-A17 core device,the open source audio library THCH-30 and aidachtang_200h are used for testing.The results show that:in the speech segment,the enhancement effect is 12dB and above;in the non-speech segment,the enhancement effect is 24dB and above,and To ensure that the voice is not obviously not distorted,compared with the rnnoise model,the processing time per 10ms is shortened by 43.29%.

关 键 词：会议通话 深度学习 消噪 非稳态噪声 

分 类 号：TN919.85[电子电信—通信与信息系统]