门控卷积神经网络多通路声学回声消除算法*

多通路声重放系统能够增强听者的现实感与空间感，但在免提通信条件下，其不可避免会受到噪声和回声干扰，严重影响通信质量。针对上述问题，本文提出了一种基于门控卷积循环神经网络的多通路声学回声消除和噪声抑制方法。该方法以传声器接收信号和重放声道的压缩复数谱为网络输入，以近端语音的压缩复数谱为网络的输出目标，直接从传声器拾取信号中恢复近端纯净语音，无需对声重放信号进行去相关处理，解决了传统自适应滤波方法中存在的非唯一解问题，同时保证了多通路声重放质量。仿真和真实声学环境实验均表明本文所提出的方法可显著消除多通路声重放系统的噪声和回声，在语音质量和回声返回衰减增益方面均优于传统算法。

Deep learning-based multichannel acoustic echo cancellation

Multichannel sound systems that utilize multi-channel audio playback devices can improve the reality and space of sound, but for hands-free communication, these systems are inevitably influenced by noise and echo, which seriously impair the communication experience. To address this issue, this paper proposes a multichannel acoustic echo cancellation and noise suppression method based on a gated convolutional recurrent neural network. This method takes the compressed complex spectrum of the near-end microphone and that of each far-end loudspeaker signal as the network input, and the compressed complex spectrum of the near-end clean speech as the network output. In this way, we can recover the clean speech from the microphone signal directly. The proposed method does not need to decorrelate the far-end signals, and thus the quality of multichannel sound reproduction is not degraded. Meanwhile, the proposed method solves the non-unique solution problem existing in the conventional adaptive filtering-based methods. Experimental results on both the simulated and real acoustic scenarios show that the proposed method can significantly suppress the noise and echo interferences in the multichannel sound system and outperform other competing methods in terms of the speech quality improvement and the echo reduction amount.