基于深神经网络的线性回波抵消与不完全传递函数的凸重构

Linear Echo Cancellation and Convex Reconstruction of Incomplete Transfer Function Based on DNN

摘要:

开发了一种在时频域中工作算法,该算法假设只有在逼近信号不活跃的频率条件下,才能估计出各自的传递函数。该算法利用在混合信号上训练的深神经网络来检测逼近信号的活动,在未检测到任何活动频率情况下,使用常规频域最小二乘法估计声波传递函数。对于出现的传递函数(ITF)估计不完整问题,该算法通过模糊时间域内ITF最稀疏表示来完成,将软阈值函数应用于时间域,由软阈值函数自适应完成,同时使用过采样来提高精度。实验结果表明：在活跃频率为80%时,该算法比传统算法收敛速度快50%左右。语音实验中,改进ADMM算法耗时0.125s,明显优于传统算法。为语音传输业务中存在回声消除问题提供了新思路。

Abstract:

An algorithm is developed in the time-frequency domain. It is assumed that the transfer function can be estimated only under the condition that the approximation signal is inactive. The algorithm uses a deep neural network trained on the mixed signal to detect the activity of the approximation signal, and the conventional frequency domain least squares method is used to estimate the acoustic wave transfer function without detecting any active frequencies. For the incomplete problem of the proposed transfer function (ITF) estimation, the algorithm is completed by the most sparse representation of the ITF in the fuzzy time domain. The soft threshold function is applied to the time domain, adaptively completed by the soft threshold function, and oversampling is used to improve the accuracy. . The experimental results show that the algorithm converges by about 50% faster than the traditional algorithm when the active frequency is 80%. In the speech experiment, the improved ADMM algorithm takes 0.125s, which is obviously superior to the traditional algorithm. It provides a new idea for the problem of echo cancellation in voice transmission services.