Microphone array signal processing for advancements in robust speech systems.

摘要

Speech system performance degrades significantly in distant-talking environments, where the speech signals can be severely distorted by additive noise and reverberation. Microphone array processing techniques have presented a potential alternative to close-talking microphones by providing speech enhancement through spatial filtering and directional discrimination. Different from conventional array optimization criteria, such as Minimal Variance Distortionless Reponse, Maximal Signal-to-Noise Ratio or Minimal Mean Squared Error, this thesis presents a series of task-oriented and environment-oriented microphone array solutions for real world speech system applications. Primarily, four important tasks (e.g., blind beamforming, automatic speech recognition (ASR), speech quality enhancement and integrated voice activity detection (VAD) with speech enhancement) are considered in two typical acoustic environments (e.g., in-vehicle and conference room). Our objective is to optimize the microphone array front-end at a system level, directly advancing the performance of a given task for the whole speech system. Specifically, several new algorithms and systems are proposed in this thesis: variance of spectra flux based blind beamforming to identify target speech source in in-vehicle and conference room environments, order statistic filter based squared spectra enhancement for ASR in in-vehicle environment, integrated VAD and speech quality enhancement system in in-vehicle environment, fast relative transfer function identification for speech quality enhancement and ASR in conference room, position dependent spectra conversion for speech quality enhancement and ASR in in-vehicle and conference room, discriminative training based VAD for in-vehicle environment, and an efficient real-time microphone array based speech acquisition platform. Primary theoretical analysis and promising real/simulation evaluations on the proposed algorithms are also presented in this thesis.