语音字典学习方法研究

Sparse coding has shown great potential in the compression, denoising, separation and classification of signals, and quickly become a hot research spot of information science. This technique represents signals over an over-complete dictionary which consists of a large number of atoms under the sparsity assumption. Therefore, the dictionary which determines the performance of dictionaries is very important.. Learning from data is an effective method to obtain atom dictionary. Although many learning methods have been proposed, most of them are originated from image processing, lacking of effect methods for speech signals. Speech signals themselves have the characteristics of weak structure, noise-like speech components, speech-like noise and complex semanteme, etc., resulting in the problems that the sparsity assumptions do not keep true completely and the performance of the dictionary is degraded. This project focuses on these problems and proposes systematically dictionary learning methods which are aimed at compression, reconstruction, denoising, separation and classification, by considering the characteristics of speech and making use of necessary prior knowledge. The goal of this project is to support speech applications and promote further improvement of speech processing techniques.

稀疏编码在信号的压缩、去噪、分离和分类等任务上表现出了巨大潜力，已成为信息科学领域的研究热点。该技术在稀疏性假定的前提下，将信号表示在由大量原子组成的过完备字典上。因此，字典在稀疏编码中至关重要，直接决定着稀疏编码的性能。. 从数据中学习是获得原子字典的有效方法。尽管目前有大量字典学习方法被提出，但绝大多数源于图像信号处理，缺乏针对语音信号的有效方法。语音信号本身具有结构性弱、类白噪成分、类语音噪声和语义复杂等特点，导致稀疏性假定不完全成立，这降低了目前方法的性能。本项目针对这些问题展开研究，充分考虑语音信号的特点，利用必要的先验知识，面向语音处理中的压缩、重构、去噪、分离和分类等任务，系统地提出字典学习的有效方法。旨在为语音相关的应用提供技术支撑，推动语音处理技术的进一步发展。

本项目在利用稀疏编码提取语音特征方面做了广泛而深入的研究，结合语音的具体特点和稀疏编码在压缩、去噪、分离和分类等任务上的优势，提出了一系列有效的方法。首先在语音字典的评价和优化方面，提出了面向信号分离和分类的评价方法和三种基于字典筛选的字典优化算法。然后在噪声字典、共结构字典学习等方面提出了字典构建方法和学习方法。最后将这些技术成功应用于语音识别任务和说话人识别任务。实验表明，所提出的方法能有效提升语音识别系统的性能，推动其走向现实应用。.项目经过四年的开展，完成了申请书里规定的目标，发表和录用论文 15 篇，其中 SCI 收录 9篇，获得省级科技进步二等奖 1 项，同时申请发明专利 6 项，获得软件著作权 4 项，发表专著 2 部。培养年轻教师2名，博士生3名，硕士生 8 人，其中已毕业硕士 2 人。.项目的开展有力地推动了语音处理领域的发展，使我们队伍掌握了智能语音的核心技术，创造了较高的经济价值和社会价值，达到了预期的目的。