Reverberation Modelling for Robust Speech Recognition in Reverberant Environments

多年来，自动语音识别（ASR）已经成功地应用于日常生活中。到目前为止，主要的限制是需要近距离交谈的麦克风，以便为自然的人/机器对话实现可接受的识别性能。然而，有许多场景，如“环境辅助生活”和“智能家居”，在这些场景中，安装在环境中固定位置的远程通话麦克风将更加方便，以便允许用户独立于麦克风位置和环境噪声进行交互。由于在这种情况下，扬声器通常离麦克风有几米远，因此接收到的信号会受到所需信号的附加噪声和混响的影响。如果不采取对策，这些影响会显著降低ASR性能。虽然在过去的几十年里，在增强ASR系统抗加性噪声方面取得了显著的进展，但混响仍然是一个主要的挑战。因此，本研究项目的关键思想是开发一个灵活的、理论上有充分根据的框架，以有效地使最先进的ASR系统适应不断变化的混响条件。在本研究项目的第一部分已经对这种方法进行了调查，现在应该进一步发展。该概念基于嵌入ASR系统的显式混响模型，旨在从先前的信号分量中估计观测信号的混响部分。混响估计器的基本结构受到混响的物理性质的启发，通过数学卷积近似，而模型参数的估计是通过利用机器学习的统计方法获得的。在该项目的前三年，随着概念的发展，在识别率方面取得了重大进展。为了进一步提高混响语音的ASR性能，项目的第二阶段侧重于将混响模型扩展到更强大的语音特征和特征向量的概率模型，因为它们主要用于最先进的ASR系统。此外，将研究不同的统计估计技术，允许仅基于少量语音信号观测对混响模型参数进行鲁棒推断。还应研究将所提出的方法与已建立的增强程序相结合，即在混响数据上训练ASR系统。由于ASR系统在实际应用中经常与麦克风阵列和信号增强算法相连接，因此最后分析了该方法与麦克风阵列信号处理概念的协同作用。

For many years, automatic speech recognition (ASR) has been successfully deployed in everyday-life applications. The main restriction so far is the necessity of close-talking microphones in order to achieve acceptable recognition performance for natural human/machine dialogues. There are, however, numerous scenarios, such as "Ambient Assisted Living" and "Smart Homes", where the employment of distant-talking microphones being installed at fixed positions in the environment would be much more convenient in order to allow the user to interact independently of the microphone positions and environmental noise. Since in such scenarios, the speaker is usually several meters away from the microphone, the received signal is impaired by additive noise and reverberation of the desired signal. These effects significantly reduce the ASR performance if no countermeasures are taken. While a remarkable progress has been achieved in robustifying ASR systems to additive noise over the past decades, reverberation still represents a major challenge. The key idea underlying this research project is, hence, to develop a flexible and theoretically well-founded framework for efficiently adapting state-of-the-art ASR systems to changing reverberation conditions. Such an approach has already been investigated during the first part of this research project and should now be further developed. The concept is based on an explicit reverberation model embedded into an ASR system, which aims at estimating the reverberant part of an observed signal from the preceding signal components. The fundamental structure of the reverberation estimator is inspired by the physical nature of reverberation approximated by a mathematical convolution, while the estimates of the model parameters are obtained by exploiting statistical methods of machine learning. During the first three years of this project, significant progress in terms of recognition rates has been achieved along with the development of the concept. In order to further the increase ASR performance for reverberant speech, the second phase of the project focuses on extending the reverberation model to more powerful speech features and probability models for the feature vectors as they are predominantly employed in state-of-the-art ASR systems. In addition, different statistical estimation techniques are to be investigated allowing for a robust inference of the reverberation model parameters based on only few speech signal observations. The combination of the proposed method with an established robustification procedure, the training of ASR systems on reverberant data, shall also be studied. As ASR systems are frequently connected to microphone arrays and signal enhancement algorithms in practical applications, the given approach is finally to be analyzed for synergies with concepts of microphone array signal processing.