基于循环神经网络的双麦克风语音增强算法

Speech Enhancement with Two Microphones Based on Recurrent Neural Network

针对基于神经网络的语音增强算法难以部署在助听器中的问题,基于循环神经网络,提出了一种低延迟、低复杂度的双麦克风语音增强算法。该算法利用两个麦克风做空域滤波初步去除非期望方向噪声,并进一步通过循环神经网络得到纯净语音信号。为了解决助听器中全相位滤波器组阶数较多而引起群延迟较大的问题,创新性地提出一种分段式滤波器组,在保证性能的同时有效减少了阶数。仿真结果显示,该滤波器组在16 k采样率下的群延迟为3.125 ms,在0 dB的babble、volvo、factory1环境下,该语音增强算法的SNR平均提升了10.556 5 dB,PESQ平均提升了0.678 7。实际测试结果中,SNR平均提升了9.439 4 dB,PESQ平均提升了0.735 0。当DSP系统时钟频率为104 MHz时,助听器经过的系统延迟为8.4 ms,功耗为6.2 mA,可以很好满足助听器高续航的需求。

Large-scale neural networks are difficult to deploy in hearing aids.A dual microphone speech enhancement algorithm based on recurrent neural networks(RNN)is proposed,which has the advantages of low latency and low complexity.The algorithm uses two microphones for beamforming to preliminarily filter out the unexpected directional noise,and further obtains the pure voice signal through RNN.In order to solve the problem of large group delay caused by excessive order of filter banks,a piecewise all phase filter bank is proposed,which can reduce the group delay and computation.The simulation results show that the group delay is 3.125 ms under 16 k sampling rate,and under the 0 dB babble,volvo,and factory1 environment,the SNR has increased by 10.556 5 dB,and the PESQ has increased by 0.678 7.In the actual test results,the SNR has increased by 9.439 4 dB,and the PESQ has increased by 0.735 0.When the clock frequency is 104 MHz,the DSP system delay of the hearing aid is 8.4 ms and the power consumption is 6.2 mA.