Blind source separation using time-frequency information obtained from wavelet analysis

利用小波分析获得的时频信息进行盲源分离

基本信息

批准号：
18540177
负责人：
MORIMOTO Akira
金额：
$ 2.34万
依托单位：
Osaka Kyoiku University
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
2006
资助国家：
日本
起止时间：
2006 至 2007
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-18540177/
关键词：
wavelet source separation time-frequency real analysis spatial mixtute snatio-tem noval mixture time delay analytic signal 時空間混合問題

项目摘要

The cocktail party effect is a challenging problem in auditory perception, which deals with the specialized human listening ability to focus one's listening attention on a single talker among a cacophony of conversations and background noise. Our interest is how to build a machine to solve the cocktail party problem in a satisfactory manner. This problem is called blind source separation. The traditional way to solve this problem is an independent component analysis. We have proposed blind source separation methods based on time-frequency information obtained from the continuous wavelet transform. There are three types of blind source separation problems, which are called spatial mixture problem, temporal mixture problem and spatio-temporal mixture problem, respectively.This research project dealt with the spatial mixture problem and the simplest type of the spatio-temporal mixture problem using wavelet analysis. Our methods have the following three merits.1. The number of sources is estimated firstly.2. Using the number of sources, the other parameters are estimated with high accuracy.3. The errors between estimated sources and real sources are small enough.We have checked these merits according to numerical simulations. Moreover, in the case of spatial mixture problem, using more than two wavelet transforms, we can estimate model parameters in a noisy environment. In the case of the simplest type of the spatio-temporal mixture problem, the source location problem using time difference of arrival measurements is considered.

鸡尾酒会效应是听觉感知中的一个具有挑战性的问题，它涉及人类专门的听力能力，即在嘈杂的对话和背景噪音中将注意力集中在单个讲话者上。我们感兴趣的是如何建造一台机器来以令人满意的方式解决鸡尾酒会问题。这个问题称为盲源分离。解决这个问题的传统方法是独立成分分析。我们提出了基于连续小波变换获得的时频信息的盲源分离方法。盲源分离问题分为三种类型，分别称为空间混合问题、时间混合问题和时空混合问题。本研究项目使用以下方法处理空间混合问题和最简单类型的时空混合问题：小波分析。我们的方法具有以下三个优点： 1．首先对来源数量进行估算； 2．利用源的数量，可以高精度地估计其他参数。 3．估计源与真实源之间的误差足够小。我们通过数值模拟检验了这些优点。此外，在空间混合问题的情况下，使用两个以上的小波变换，我们可以在噪声环境中估计模型参数。在最简单类型的时空混合问题中，考虑使用到达测量时间差的源定位问题。