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

发布时间：2018-06-13 02:08

本文选题：深度神经网络 + 语音增强　；参考：《计算机应用研究》2017年03期

【摘要】：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,该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。
[Abstract]:Deep neural Networks (DNNs) has been widely used in speech enhancement task because of its good feature extraction ability. In order to improve the speech enhancement effect of depth neural network, a new network structure is proposed, which combines depth neural network and constrained Wiener filter. The network firstly trains the noisy speech amplitude spectrum and obtains the pure speech and noise amplitude spectrum estimation, and then calculates a constrained Wiener gain function by using the speech and noise amplitude spectrum estimation. Finally, the constrained Wiener gain function is used to estimate the enhanced speech amplitude spectrum from the noisy speech amplitude spectrum as the training output of the network. The simulation results of 20 kinds of noise under different SNR show that the proposed method can effectively remove the noise and keep the speech distortion small regardless of whether the noise type appears in the training set of the network. The enhancement effect is better than that of DNN and NMF.
【作者单位】：解放军理工大学指挥信息系统学院;东部战区综合信息服务队;西安通信学院;
【基金】：国家自然科学基金资助项目(61471394,61402519) 江苏省自然科学基金资助项目(BK20140071,BK20140074)
【分类号】：TP183;TN912.35

【相似文献】