联合优化深度神经网络和约束维纳滤波的单通道语音增强方法

深度神经网络（Deep neural networks，DNNs）依靠其良好的特征提取能力，在语音增强任务中得到了广泛应用。为进一步提高深度神经网络的语音增强效果，提出一种将深度神经网络和约束维纳滤波联合训练优化的新型网络结构。该网络首先对带噪语音幅度谱进行训练并分别得到纯净语音和噪声的幅度谱估计，然后利用语音和噪声的幅度谱估计计算得到一个约束维纳增益函数，最后利用约束维纳增益函数从带噪语音幅度谱中估计出增强语音幅度谱作为网络的训练输出。对不同信噪比下的20种噪声进行的仿真实验表明，无论噪声类型是否在网络的训练集中出现，本文方法都能够在有效去除噪声的同时保持较小的语音失真，增强效果明显优于DNN及NMF增强方法。

Joint optimization of Deep Neural Networks and constrained Wiener Filter for single channel speech enhancement

Depending on its higher ability of extracting features, deep neural networks are widely used in speech enhancement. To improve the speech enhancement performance of DNN, this paper proposed a new DNN architecture which joint optimization with constrained wiener filter. Firstly, the proposed neural network directly trained the noisy magnitude spectrum and got both the clean speech magnitude spectrum estimator and the noise magnitude spectrum estimator. Then, the network combined clean speech magnitude spectrum estimator and the noise magnitude spectrum estimator to calculate a constrained wiener gain function. Finally, the network obtained the enhanced speech magnitude spectrum as the DNN output from the noisy speech magnitude spectrum by using constrained wiener gain function. Experiments with 20 noise types at various SNR levels demonstrate that the proposed method outperforms the DNN method and the NMF method, which can effectively removes the noise while maintaining smaller speech distortion, no matter whether the noise conditions are included in the training set or not.