Existing deep learning-based speech denoising approaches require clean speech signals to be available for training. This paper presents a deep learning-based approach to improve speech denoising in real-world audio environments by not requiring the availability of clean speech signals as reference in training mode. A fully convolutional neural network is trained by using two noisy realizations of the same speech signal, one used as the input and the other as the target of the network. Two noisy realizations of the same speech signal are generated by using a mid-side stereo microphone. Extensive experimentations are conducted to show the superiority of the developed deep speech denoising approach over the conventional supervised deep speech denoising approach based on four commonly used performance metrics as well as a subjective testing.

现有的基于深度学习的语音去噪方法要求干净的语音信号可用于训练。本文提出了一种基于深度学习的方法，通过在训练模式下不要求提供干净的语音信号作为参考来改善现实世界音频环境中的语音降噪。通过使用同一语音信号的两个有噪实现来训练全卷积神经网络，一个用作输入，另一个用作网络的目标。通过使用中侧立体声麦克风，可以产生相同语音信号的两个噪声实现。进行了广泛的实验，以显示开发的深度语音去噪方法相对于基于四个常用性能指标以及主观测试的传统监督式深度语音去噪方法的优越性。