困扰语音识别走向实用的羁绊之一是噪声消除问题，而传统的语音增强方法难以应对非平稳噪音，如:人声背景噪声，计算听觉场景分析（CASA）的语音分离方法提供了一条颇具潜力的解决途径。现有CASA系统大致分为基于数据驱动和基于模式驱动两种方式，数据驱动偏重强调利用底层声学线索的自下而上的处理机制；而模式驱动则强调利用人的学习机制的自上而下的处理方式，迄今缺少有效融合两种方式的CASA系统。事实上，人的听觉感知系统是两种方式并存的有机整体，基于此，本项研究提出基于数据和模式双驱动的CASA语音分离新方法，旨在克服现有系统单一驱动的缺陷，使CASA技术获得跨越发展。主要研究内容包括：数据驱动核心基础的多基频提取和组织方法、模式驱动关键环节的学习和分类模型、以及数据和模式双驱动的有效融合方法等。本项研究对于数字媒体领域的多媒体检索、声学目标检索、语音增强、鲁棒性语音识别等关键技术具有重要意义和学术价值。

One of fetters to perplex speech recognition applications is background noise cancelation while conventional speech enhancement algorithms are hard to handle unstable noise, for example, background noise of speech, however speech seperation approach with computational auditory scene analysis(CASA) provides a well potential way. The available CASA systems are divided roughly into two categories, namely data driven based and schema driven based, to which the former emphasezes bottem up mechanism by utilizing primitive acoustic cues, while the latter stresses top dowm processing way of human learning ability. Up to now it is short of CASA system to cooperate two kinds of processing mechanism efficiently. Base on the fact that human auditory perception system is an organic integrity with the existence of both two processing ways, a CASA novel methed of speech separation is presented based on both data driven and schema driven in the proposal.The goal is to conquer the flaw of single driven style in the available CASA system and promotes CASA technology leapfrog development. The main research contents include kernel groundwork of data driven policy such as multipitch extraction and grouping strategy in data driven, crucial points of schema driven way such as machine learning or classifying model, and efficient fusion algorithm for both data driven and schema driven,and so on . The reseach is very important and with academic value for the key technologies in the field of digital media such as multimedia retrieval, acoustic target retrieval, speech enhancement and robust speech recognition etc.