一种基于组合深层模型的语音增强方法

Integration of Deep Networks for Speech Enhancement

模型建立是语音增强系统的重要一环,对最终系统性能起着决定性的作用。针对语音增强系统在低信噪比和非平稳噪声环境中性能不佳的问题,提出一种基于组合深层模型的语音增强系统。首先,使用深度神经网络(deep neural network,DNN)从含噪语料中估计掩蔽值;然后,将掩蔽值通过前后帧拼接转化为灰度图送入卷积神经网络(convolutional neural network,CNN)进行识别;最后,用识别出的掩蔽矩阵合成目标语音。实验证明,与单纯使用DNN进行掩蔽值估计的系统相比,联合DNN和CNN的语音增强系统在各项评价指标中均得到改进,尤其在低信噪比和非平稳噪声环境中提升更为明显。

Network plays a vital role in the speech enhancement system. In this paper, a novel speech enhancement system based on the integration of deep networks is proposed to solve the problem that the performance of speech enhancement system is poor in low signal-to-noise ratio (SNR) and non-stationary noise environments. Firstly, the deep neural network (DNN) is used to estimate mask values from the noisy utterances. Then mask values are mapped to a gray-scale map by catenating the neighboring frames. Afterwards, the convolutional neural network (CNN) is applied to recognize the mask matrix. Finally, the recognized mask matrix is used to synthesize the target speech. Experiment results show that compared to DNN based speech enhancement system, the integration of DNN and CNN achieves improvements in terms of segmental signal-to-noise ratio, perceptual evaluation of speech quality and short-time objective intelligibility, especially in low SNR and non-stationary noise environments.