自回归维纳滤波语音增强方法研究

项目名称： 自回归维纳滤波语音增强方法研究

项目编号： No.61471014

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 鲍长春

作者单位： 北京工业大学

项目金额： 80万元

中文摘要： 本课题以目前移动通信中普遍使用的线性预测语音编码技术为背景，通过构建能量依赖的有限状态语音谱包络先验码书，研究复杂环境下的自回归（AR）维纳滤波语音增强方法。该方法将利用极大似然估计确定最终的语音AR谱包络，并借助估计的噪声谱和观测信号谱之间的相关性修正维纳滤波器，以达到平衡语音和噪声功率的目的。另外，该方法还将对观测信号进行相空间重构来改善噪声估计性能，并用深度信念网络和模拟退火等算法优化语音的AR谱包络码书。该方法不需噪声分类，非常适合移动通信中噪声类型和噪声能量逐帧改变的非平稳噪声抑制。

中文关键词： 语音增强；语音编码；线性预测；自回归模型；维纳滤波

英文摘要： Considering the background of linear prediction speech coding that has been widely used in mobile communication, the speech enhancement method in adverse environment based on auto-regressive (AR) Wiener filtering (WF) will be proposed in this project. This method exploits the priori codebooks about spectral shape of speech to optimize AR power spectrum shape of the speech based on maximum-likelihood estimation, where the priori codebooks have finite states that depend on the energy of speech. This method also tries to find the statistical cross-correlation between the spectra of noise and noisy observation to modify Wiener filter for balancing the power ratio of speech and noise. In addition, the noise estimation will be improved by phase space reconstruction of the noisy observation, and the priori codebooks about spectral shape of speech will be optimized by Deep Belief Network(DBN) and simulated annealing(SA) algorithm.This kind of speech enhancement method can avoid the noise classification, and is very suitable for suppressing nonstationary noise that the types and energy vary on a frame-by-frame basis in mobile speech communication.

英文关键词： Speech enhancement;Speech coding;Linear prediction;Auto-regressive model;Wiener filtering