Single and multi-microphone techniques of speech enhancement for reliable voice communications under adverse acoustic conditions

针对各种环境声噪声的语音增强是许多语音处理设备和系统的基本组成部分。它发现了广泛的应用，从用于无线接入的移动电话和pda，到用于低成本互联网接入的软件定义工具。降噪和语音增强算法不仅有助于提高语音信号的质量，而且还可以作为前端处理来提高语音编码和识别等其他应用的性能。尽管近年来取得了许多进展，但受中等至高水平非平稳噪声（例如：自助餐厅、繁忙的街道、体育赛事）污染的语音增强仍然是一个具有挑战性的研究途径，也是我们赞助商关注的一个具体问题。因此，该项目的主要目标是为这一关键任务开发改进的信号处理算法。建议的工作包括沿着两个主要途径构建的几个子项目，即：(1)开发新的单通道语音增强核心技术；(2)将这些技术扩展到配备多麦克风的音频设备。第一个主题旨在探索新的基于码本的方法，利用关于说话人和背景噪声的可用信息，并研究使用先锋压缩感知原理来减少非平稳噪声。第二个主题侧重于开发新的波束形成技术，该技术结合压缩感知，以实现多个麦克风信号的最佳相干组合，并研究这些系统所需的噪声场统计估计的新方法。该项目中提出的改进语音增强算法将为我们的赞助商Microsemi提供实用且具有成本效益的解决方案，以应对非平稳噪声带来的许多挑战，以开发和商业化用于语音应用的新DSP产品系列，从而帮助公司扩大其客户群和年销售额。该项目还包括一个重要的培训部分，将在麦吉尔大学和康考迪亚大学进行。

Speech enhancement against various environmental acoustic noises is a fundamental component of many voice processing devices and systems. It finds wide applications ranging from mobile phones and PDAs for wireless access, to software-defined tools for low-cost internet-based access. Noise reduction and speech enhancement algorithms not only contribute to improve the quality of speech signals, but also serve as front-end processing to improve the performance of other applications such as speech coding and recognition. In spite of many advances in recent years, the enhancement of speech contaminated by moderate to high levels of non-stationary noise (e.g: cafeteria, busy street, sporting events) remains a challenging research avenue and a specific problem of concern to our sponsor. The main objective of this project is therefore to develop improved signal processing algorithms for this critical task. The proposed work consists of several sub-projects structured along two main avenues, namely: (1) development of new core techniques for single-channel speech enhancement and; (2) extension of these techniques for audio devices equipped with multi-microphones. The first topic aims to explore new codebook-based approaches that take advantage of available information about the speaker and background noise, and to investigate the use of vanguard compressive sensing principles for the reduction of non-stationary noise. The second topic focuses on the development of new beamforming techniques that incorporate compressive sensing to enable optimal coherent combination of the multiple microphone signals, and the investigation of new approaches for estimation of the noise field statistics, as required in these systems. The improved speech enhancement algorithms proposed in this project will provide our sponsor, Microsemi with practical and cost-effective solutions to the many challenges posed by non-stationary acoustic noise in its quest to develop and commercialize new lines of DSP products for voice applications, thereby helping the company to grow its customer base and annual sales. The project also includes a significant training component that will take place at both McGill and Concordia universities.