Microphone array processing techniques for the enhancement, dereverberation and separation of speech

用于通过无线和互联网接入进行语音和多媒体通信的消费产品，例如手机、手持平板电脑、VoIP电话等，广泛使用数字语音处理（DSP）来传输、存储和回放信息。语音增强作为语音编码或语音识别的核心技术和前端，在这些设备中起着关键的作用。实际上，随着它们的使用变得广泛，由麦克风接收的语音信号被越来越不利的声学干扰（包括混响和非平稳噪声）劣化，这降低了期望语音的质量和可懂度。因此，强烈需要改进的增强技术，其可以抑制这些声学干扰并产生用于传输的更清晰的语音信号。在这个项目中，我们将研究新的DSP技术，用于增强混响语音及其从非平稳声学背景中分离。本文的主要工作包括：（1）将多通道短时谱幅度（STSA）语音估计器扩展到混响环境;（2）基于子带分解的联合去混响降噪技术;（3）非负矩阵分解（NMF）语音分离算法研究;（4）系统集成和麦克风阵列验证。这项拟议的研究是在工业合作伙伴Microsemi的赞助下成功完成的NSERC CRD项目的延伸。尽管该公司的宽带语音处理平台越来越受欢迎，但非平稳噪声的减少、语音的去混响及其与声学背景的分离仍然是阻碍进一步技术进步的挑战性问题。拟议的研究旨在克服这些限制，并为Microsemi提供创新的，具有成本效益的解决方案来应对这些挑战。除了技术转让，该项目将有助于推进语音和音频处理的工程学科，并在麦吉尔大学和康考迪亚大学培训这一领域的高素质人才。

Consumer products for voice and multimedia communications over wireless and internet access, e.g. cell phones, handheld tablets, VoIP phones, etc., make extensive use of digital speech processing (DSP) for the transmission, storage and playback of information. Speech enhancement, as a core technology and front-end to speech coding or automatic speech recognition, plays a key role in these devices. Indeed, as their usage becomes widespread, the voice signals received by the microphones are degraded by ever more adverse acoustic disturbances, including reverberation and non-stationary noise, which degrade the quality and intelligibility of the desired speech. Hence, there is a strong demand for improved enhancement techniques that can suppress these acoustic disturbances and produce a cleaner speech signal for transmission. In this project, we will investigate new DSP techniques for the enhancement of reverberant speech and its separation from a non-stationary acoustic background. Our work aims to achieve the following objectives: (1) Extension of multi-channel short time spectral amplitude (STSA) speech estimators to reverberant environments; (2) Development of joint dereverberation and noise reduction techniques based on subband decomposition; (3) Investigation of non-negative matrix factorization (NMF) algorithms for speech separation; and (4) System integration and validation with microphone array. This proposed research is an extension of a successfully completed NSERC CRD project under the sponsorship of industrial partner, Microsemi. In spite of the growing popularity of the company's broadband voice processing platforms, the reduction of non-stationary noise, the dereverberation of speech and its separation from acoustic backgrounds, remain challenging problems that hinder further technological progress. The proposed research aims to push back such limitations and provide Microsemi with innovative, cost-effective solutions to these challenges. In addition to technology transfer, the project will contribute to advance the engineering discipline of speech and audio processing and as to the training of highly qualified personal in this area at McGill and Concordia Universities.