基于EEMD数据预处理和DNN的语音增强算法研究

Research on Speech Enhancement Algorithm Based on EEMD Data Preprocessing and DNN

提出了基于总体平均经验模态分解(EEMD)预处理和深度神经网络的语音增强算法,首先将带噪语音信号和纯净语音信号进行EEMD分解,获得一组频率从高到低的本征模态函数IMF分量,然后从各IMF中提取时域的信号特征,组成特征向量,输入神经网络中进行训练。实验表明:该算法与经典无监督算法比,无需任何假设条件,可以较好地学习带噪语音和纯净语音之间复杂的非线性关系,在语音质量和可懂度方面优势明显,显示了深度神经网络在语音增强方面的独特作用。

A speech enhancement algorithm based on EEMD(Ensemble Empirical Mode Decomposition)preprocessing and deep neural network was proposed.The noisy speech signal and the pure speech signal were decomposed by EEMD to get a set of IMF(Intrinsic Mode Function)components from high to low frequency.Then,the signal features in the time domain are extracted from each IMF,and the feature vectors were constructed,they are input into the neural network for training.Experiments show that the algorithm does not require any assumptions when compared with the classical unsupervised algorithm.It can better learn the complex nonlinear relationship between noisy speech and pure speech.It has obvious advantages in speech quality and intelligibility,showing the unique role of deep neural network in speech enhancement.