Mitigation of noise and reverberation for hearing aids, voice-enabled devices, and speech transmission

这项建议是为了开发改进的去噪算法，以减少加性噪声和混响在以下三个应用中造成的损害，这三个应用对加拿大经济都非常重要。1)助听器：随着加拿大65岁以上人口的比例越来越高，助听器变得越来越重要。2)支持语音的设备：通常使用免提语音技术，它们正在迅速崛起。3)语音传输：仍然是一个热门话题(推出IP语音服务等)，仍然需要开发更有效的预处理器，以便在语音编码之前从语音中去除附加噪声或混响。在我们的团队中，我们最近开发了一种在加性噪声中进行语音增强的方法，该方法将人类的听觉特征整合到强大的统计建模和跟踪算法中。在提出的方案中，我们试图使这些语音增强算法适用于助听器的去噪问题。我们还希望研究我们的语音增强算法在支持语音的设备环境中的效率。我们最近开发了一些很有前途的盲源分离(BSS)算法，并打算将它们应用于语音中的加性噪声去除。由于从语音信号中去除混响效应的问题可以归结为盲解卷积问题，我们还打算将本课题组开发的有前途的盲源分离算法扩展到卷积混合的情况，提供一种有效的去混响方法。我们还试图将听觉建模方面包含到由此产生的盲解卷积算法中。此外，还将研究使用强大的和新兴的粒子滤波方法来抑制语音中的加性噪声。粒子滤波器最近被用于盲解卷积，因此我们还将研究它们在语音去混响方面的应用。调查将听觉模型或感知方面纳入粒子过滤器训练的不同方法是该计划的另一个目标。最后，我们还将寻求在语音去噪的背景下降低粒子过滤器的复杂性的方法。

This proposal is for the development of improved de-noising algorithms to reduce the impairments caused by additive noise and reverberation in the following three applications, which are all of significant importance for the Canadian economy. 1) Hearing aids: they are becoming increasingly important as the proportion of Canadians over 65 is getting higher. 2)Voice enabled devices: typically using hands-free speech technology, they are quickly emerging. 3)Speech transmission: still a hot topic (launch of Voice over IP services, etc.), and there is still a need for developing more efficient pre-processors to remove additive noise or reverberation from speech before it is encoded. In our group we have recently developed an approach for speech enhancement in additive noise, which integrates human hearing characteristics into powerful statistical modeling and tracking algorithms. In the proposed program we seek to adapt these speech enhancement algorithms to the problem of de-noising for hearing aids. We also want to investigate the efficiency of our speech enhancement algorithms in the context of voice enabled devices. We have recently developed some promising Blind Source Separation (BSS) algorithms, and we intend to apply and test them for removing additive noise in speech. Since the problem of removing a reverberation effect from a speech signal can be formulated as a problem of Blind Deconvolution, we also intend to extend the promising BSS algorithms developed in our group to the case of convolutive mixtures, providing a powerful method to perform de-reverberation. We also seek to include auditory modeling aspects into the resulting blind deconvolution algorithms. Moreover, the use of the powerful and emerging particle filter methods for the mitigation of additive noise in speech will be investigated. Particle filters have recently been used for blind deconvolution, so we will also investigate their use in the context of speech de-reverberation. The investigation of different ways of including auditory models or perceptual aspects into the training of particle filters is another objective of the program. Finally, we will also seek to find ways to reduce the complexity of particle filters in the context of speech de-noising.