基于树莓派的智能语音降噪算法研究与实现  

Research and Implementation of Smart Speech Noise Reduction Algorithm Based on Raspberry Pi

作　　者：陶然[1] 朱润乾 秦怡童 凌海东 TAO Ran;ZHU Runqian;QIN Yitong;LING Haidong(School of Communication and Artificial Intelligence,School of Integrated Circuits,Nanjing Institute of Technology,Nanjing 211167,China)

机构地区：[1]南京工程学院通信与人工智能学院、集成电路学院,江苏南京211167

出　　处：《现代信息科技》2025年第3期183-188,共6页Modern Information Technology

摘　　要：语音增强是语音信号处理的重要分支,在语音识别、语音通信等领域具有重要应用。传统数字信号处理(DSP)方法下的单通道语音增强计算量小,但效果不佳。近年来,深度学习算法大幅提升了单通道语音增强的效果,但往往计算量大,对硬件要求高,难以应用于移动设备或可穿戴设备。针对性能和计算量难以平衡的现状,文章实现了一种低复杂度的基于深度学习的语音增强算法,并在树莓派上进行了实现。该算法采用具有四个隐藏层的循环神经网络(RNN),用于估计理想的临界频带增益,而音高谐波之间的噪声则采用传统音高滤波器处理。实验结果显示,该系统能够有效实现降噪功能,并且性能优于传统的维纳滤波算法。Speech enhancement is an important branch of speech signal processing and has significant applications in fields such as speech recognition and speech communication.The single-channel speech enhancement under traditional Digital Signal Processing(DSP)method has a small amount of computation,but the effect is not satisfactory.In recent years,Deep Learning algorithms have significantly improved the effect of single-channel speech enhancement.However,they usually have a large amount of computation and high hardware requirements,making it difficult to apply them to mobile or wearable devices.In view of the current situation where it is difficult to balance performance and computation,this paper implements a low-complexity Deep Learning-based speech enhancement algorithm and realizes it on a Raspberry Pi.This algorithm adopts a Recurrent Neural Network(RNN)with four hidden layers to estimate the ideal critical band gain,while the noise between pitch harmonics is processed using traditional pitch filters.Experimental results show that this system can effectively achieve the noise reduction function and outperforms the traditional Wiener filtering algorithm.

关 键 词：语音增强 RNNoise 实时 单通道 树莓派 

分 类 号：TN912[电子电信—通信与信息系统]