项目名称: 基于压缩感知的单通道混合语音分离理论及算法研究
项目编号: No.61302152
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 无线电电子学、电信技术
项目作者: 郭海燕
作者单位: 东南大学
项目金额: 24万元
中文摘要: 一直以来,单通道混合语音分离都是语音信号处理的重要研究方向。近几年来出现的研究热点—压缩感知,为单通道混合语音分离的研究提供了新的思路。本项目先基于压缩感知,研究单通道混合语音准确分离的理论条件,包括:新的基于压缩感知的单通道混合语音分离建模方法,准确分离和准确重构的等价条件分析和准确重构条件的理论分析。在此理论基础上,本项目还研究设计实用的基于压缩感知的单通道混合语音分离算法,包括:适用于小规模训练数据的语音信号自适应稀疏基的构造算法,基于改进阶梯正交匹配追踪的快速分离算法,和针对含清音混合语音帧的双重分离方案。本项目的研究成果可广泛用于语音增强、鲁棒语音识别、鲁棒说话人识别、电话会议、助听器设计等领域。
中文关键词: 语音分离;压缩感知;稀疏分解;字典学习;
英文摘要: Compressed sensing (CS), emerged in recent years, can provide a new way to solve the problem of single-channel Speech Separation(SCSS), which is a vital issue of speech signal processing. Therefore, we put our research focuses on CS-based SCSS techniques. The research includes two parts. One is to discuss the conditions of exact CS-based separation, including new CS-based SCSS modelling, the analysis of conditions in which the exact separation problem is equivant to the exact recovery problem and the analysis of exact recovery conditions.The other is to design practical CS-based separation methods, including an adaptive basis construction algorithm suitable for small-scale training data,a fast CS-SCSS algorithm based on improved stagewise orthogonal matching pursuit and a double separation scheme for unvoied/voiced speech mixture. Our achievements in this program can be widely applied to speech enhancement, robust speech recognition, robust speaker recognition, teleconference, hearing aid design and so on.
英文关键词: speech separation;compressed sensing;sparse decomposition;dictionary learning;