本文为英国伦敦帝国理工学院（作者：Clement S. J. Doire）的博士毕业论文，共177页。

当在有限的声学空间内使用远距离麦克风捕捉语音信号时，录音往往会因混响而退化。这会对语音的质量和可理解性产生有害影响，尤其是当与噪声混合时。近几年来，在免提电话或听音器技术等应用中，人们越来越需要有效的方法来对抗混响的破坏性影响。然而，如何提供一种对高噪声具有鲁棒性且适合实时处理的盲单通道去冗余方法仍然是一个挑战。许多单通道去混响算法的一个重要前提是估计控制混响的声学参数。在本论文中，我们提出了一种新的结合干扰信号功率来估计这些参数的新方法，该方法是基于语音活动检测和扩展卡尔曼滤波器相结合的。然后将该方法扩展到考虑干净语音信号的频谱结构，并通过对退化语音谱图应用时频增益来进行去冗余处理。该增益的估计被描述为一个基于隐马尔可夫模型的贝叶斯滤波问题。为了从语音清晰度的角度评价该算法，提出了一种在听力实验中有效测量心理功能的新算法。所开发的算法将在模拟和真实记录上进行评估，并与现有最先进的替代方案进行比较。

When capturing speech signals using adistant microphone within a confined acoustic space, the recordings are oftendegraded by reverberation. This can have a detrimental impact on the qualityand intelligibility of speech, especially when combined with acoustic noise. Inrecent years, there has been increasing demand for effective ways of combatingthe damaging effects of reverberation in applications such as hands-freetelephony or hearingaids technology. However, the task of providing a blindsingle-channel dereverberation method robust to high levels of noise andsuitable for real-time processing remains a challenge. An importantprerequisite for many single-channel dereverberation algorithms is theestimation of the acoustic parameters governing reverberation. In this thesis, anovel online method of estimating these parameters jointly with the interferingsignal powers is proposed that is based on a combination of Voice ActivityDetection and Extended Kalman Filters. This method is then extended to takeinto account the spectral structure of clean speech signals and to performdereverberation by applying a time-frequency gain to the degraded speechspectrogram. The estimation of this gain is formulated as a Bayesian filteringproblem conditioned on a Hidden Markov Model. In order to evaluate the proposedalgorithm in terms of speech intelligibility, a novel algorithm for measuringPsychometric Functions efficiently in listening experiments is presented. Thealgorithms developed are evaluated on both simulated and real recordings andare compared with existing state-of-the art alternatives.

1. 引言

2. 项目背景与文献回顾

3. 混响参数的盲估计算法

4. 单通道语音增强技术

5. 有效的语音清晰度估计

6. 结论

更多精彩文章请关注公众号：