Deep neural network-based speech enhancement for robust speech recognition in smart home device

基于自然语音的人机界面已经发展到了以前不可想象的应用程序正在成为我们日常生活的一部分的阶段。语音接口不仅促进了人机交互，还显著提高了家庭自动化的效率，这是物联网（IoT）的关键驱动力。智能家居设备（SHD）现在已经上市，允许用户远程控制他们的家庭设备并访问基于Web的信息源。这些智能助手可以通过自动语音识别（ASR）实时响应人类语音命令。然而，SHD令人满意地运行在现实世界的条件下，他们必须是强大的声学噪声和混响，一个关键的问题，其解决方案需要新的语音处理技术。该项目的长期目标是开发一个基于深度神经网络（DNN）的集成语音增强（SE）系统，以支持两个基本的SHD功能：关键字识别和基于云的ASR。在为期3年的时间里，该项目旨在实现以下目标：开发用于表示嘈杂语音的新功能集;设计更适合SE任务的改进的DNN核心引擎;实现一个完整的基于DNN的SE系统，该系统对噪声和混响具有鲁棒性;最后，在多麦克风SHD环境中评估其性能。这项拟议的研究是与工业合作伙伴Microsemi正在进行的NSERC CRD项目的延伸。在过去的两年中，我们的团队开发了最先进的SE算法，在人类听众身上测试时表现出色。然而，这些算法并没有被最佳地设计为用作ASR的预处理器，而这正是Microsemi目前正在开发的用于语音驱动SHD的新系列集成电路（IC）所需要的。拟议的研究将为我们的赞助商提供具有成本效益和创新的SE解决方案，用于其IC产品，提高其在市场上的竞争力。除技术转让外，该项目还将促进麦吉尔大学和康考迪亚大学在智能语音处理方面的研究和HQP培训。

Human-machine interfaces based on natural speech have advanced to the stage where previously unthinkable applications are becoming part of our daily life. Speech interfaces not only facilitate human-machine interactions but also significantly enhance the efficiency of home automation, which is a key driver of internet of things (IoT). Smart home devices (SHD) are now commercially available that allow users to control their home gears remotely and access Web-based information sources. These intelligent assistants can respond in real-time to human voice commands via automatic speech recognition (ASR). However, for SHD to operate satisfactorily under real-world conditions, they must be robust to acoustic noise and reverberation, a critical problem whose solution calls for new speech processing technologies. The long-term goal of the project is to develop an integrated speech enhancement (SE) system based on deep neural networks (DNN) to support two essential SHD functions: keyword spotting and cloud-based ASR. Over its 3-year duration, the project aims to achieve the following objectives: develop new feature sets for the representation of noisy speech; design improved DNN core engines better suited to the SE task; implement a complete DNN-based SE system that is robust to noise and reverberation; and finally, evaluate its performance within a multi-microphone SHD context. This proposed research is an extension of an on-going NSERC CRD project with industrial partner Microsemi. During the past two years, our team has developed state-of-the-art SE algorithms that display excellent performance when tested on human listeners. However, these algorithms are not optimally designed for use as pre-processors to ASR, as needed in the new line of integrated circuits (IC) being currently developed by Microsemi for voice-driven SHD. The proposed research will provide our sponsor with cost-effective and innovative SE solutions for use in their IC products, boosting its competitiveness on the marketplace. In addition to technology transfer, the project will promote the research and the training of HQP in intelligent speech processing at McGill and Concordia.************