Deep Learning Technologies for Acoustic Echo Cancellation in Dynamic Environments

在全双工免提语音通信系统中，位于链路一端的房间中的扬声器可能由于扬声器与链路另一端的麦克风之间的声学耦合而接收到他/她的语音的回声。声学回声消除（AEC）的目标是消除这种不期望的回声，以提高语音通信的质量和可懂度，以及相关应用（例如自动语音识别（ASR））的性能。虽然过去已经提出了许多基于传统自适应滤波技术的AEC算法，但最近关于深度神经网络（DNN）应用于该问题的研究已经显示出显着的性能，特别是在不利条件下，即，然而，这些DNN推广到声学环境的广泛动态的有限能力仍然是一个开放的研究问题，也是我们的合作伙伴Fluent.ai关注的具体问题，其重点是开发下一代语音用户界面（VUI）。在此框架内，该项目的主要目标是开发新的基于DNN的AEC算法，以克服实时应用的上述限制。我们建议的工作包括：（1）将由声学脉冲响应的在线估计提供的附加信息合并到基于DNN的AEC框架中，以及（2）确定用于实现该方法的最合适的DNN架构。该项目开发的新算法将被Fluent.ai用于建立并最终商业化用于语音通信和ASR的嵌入式VUI新产品线，从而使公司能够吸引更多客户并发展业务。除了满足公司的具体需求外，这个短期项目还将促进麦吉尔大学的学术研究人员与Fluent.ai的科学成员之间建立新的研究伙伴关系，为加拿大带来重大的长期利益。

In a full-duplex hands-free voice communication system, a speaker located in a room at one end of the link may receive an echo of his/her voice due to the acoustic coupling between the loudspeaker and the microphone at the other end of the link. The goal of acoustic echo cancellation (AEC) is to remove such undesirable echo in order to improve the quality and intelligibility of the voice communication, as well as the performance of related applications, such as automatic speech recognition (ASR). While numerous AEC algorithms based on traditional adaptive filtering techniques have been proposed in the past, recent studies on the application of deep neural networks (DNN) to this problem have shown remarkable performance, especially under adverse conditions, i.e., high levels of noise and reverberation, nonlinear characteristics of audio devices, etc. However, the limited ability of these DNN to generalize to the wide dynamics of the acoustic environment still remains an open issue for research and a specific problem of concern to our partner, Fluent.ai, whose focus is on developing the next generation of voice user interfaces (VUI). Within this framework, the main objective of this project is to develop new DNN-based AEC algorithms to overcome the above limitation for real-time applications. Our proposed work includes: (1) incorporating additional information provided by on-line estimation of the acoustic impulse response into the DNN-based AEC framework, and (2) determining the most suitable DNN architecture for implementing this approach. The new algorithms developed in this project will be used by Fluent.ai to establish and eventually commercialize a new line of embedded VUI for voice communications and ASR, thereby enabling the company to attract additional customers and grow its business. In addition to address the company's specific need, this short-term project will foster the development of a new research partnership between the academic researchers at McGill University and the scientific members at Fluent.ai, bringing significant long-term benefits to Canada.