Model-based Analysis-by-Synthesis for the Dereverberation of Speech and Audio Signals

该项目通过多通道盲通道识别和均衡来处理语音信号的去混响。这项最后提案旨在通过使用统计方法系统地改进目前的方法。即将到来的优化将基于贝叶斯推理原理，将现有的最大似然(ML)解和最大后验(MAP)估计器扩展为变分贝叶斯(VB)估计算法。对于未知的声学系统和源信号，较强的模型允许推导出强大的VB算法。基于现有的用于信道的马尔可夫模型，最后一个项目阶段将特别解决源信号的建模。通过引入关于未知量的先验信息，所得到的算法将专门针对去混响问题而定制，以便减少剩余的估计误差。VB意义上的优化最终应该导致一种有效的估计算法，该算法自适应地估计信道和源信号以及相应的模型参数。此外，对拟议的系统从ML到MAP和VB的系统改进还允许对所有三种估计方法进行结构性比较，并分析其各自的优点和缺点。所有这些都将有助于评估拟议方法的整体效益。

This project deals with the dereverberation of speech signals by means of multichannel blind channel identification and equalization. This final proposal aims at a systematic enhancement of the current approach by using statistical methods. The upcoming optimization will be based on Bayesian inference principles that extend the existing Maximum Likelihood (ML) solution and a Maximum A Posterior (MAP) estimator towards a Variational Bayes (VB) estimation algorithm.Strong models for the unknown acoustic system and the source signal allow for the derivation of powerful VB algorithms. Based upon the existing Markov model for the channels, this last project phase will particularly address the modeling of the source signal. By introducing a priori information about the unknown quantity, the resulting algorithms will then be specifically tailored to the dereverberation problem in order to mitigate remaining estimation errors. The optimization in a VB sense should eventually lead to an efficient estimation algorithm that adaptively estimates the channel and the source signal as well as the corresponding model parameters. Such a systematic enhancement of the proposed system from ML to MAP and VB moreover allows for a structural comparison of all three estimation approaches and an analysis of their individual strengths and shortcomings. All this will help to assess the overall benefit of the proposed methods.