The need for separating independent speech signals using multiple microphones in a reverberant environment arises in a variety of applications; e.g., speech enhancement, speech recognition, hands-free telephony, etc. In most of these applications, very little or nothing is known about the source signals or the way they are mixed together, making the separation methods “blind.” Existing blind source separation (BSS) methods tend to break down in a realistic reverberant environment. In this thesis, we show that this limited performance is due to random permutations of the unmixing filters over frequency. We refer to this problem as permutation inconsistency, which becomes worse as the length of the room impulse response increases. By developing diagnostic tools, we reveal that if the unmixing filter matrix permutations are properly aligned at all frequency bins, the performance of the BSS method is greatly improved. We derive ideal separation performance benchmarks and examine the effect of microphone separation and room reverberation on the separation performance.; We study the performance of an ideal null-steering beamformer in the context of speech separation when the source locations are assumed to be known. This leads us to explore interesting connections between BSS and ideal beamforming, where we show the feasibility of using beamformer concepts to resolve the permutation inconsistency problem. We propose a permutation alignment scheme based on information gathered from microphone array directivity patterns. This technique is novel in the sense that it works satisfactorily even when the directivity patterns exhibit grating lobes. We also illustrate the remarkable performance of BSS, which outshines the ideal beamformer in highly-reverberant environments even though the later assumes a prior knowledge of speech source locations.; Finally, we discover the phenomenon of the loss of spectral resolution when one tries to align the unmixing filter permutations. We refer to this conflict between the two requirements as permutation-inconsistency /spectral-resolution tradeoff. To ease this tradeoff, we propose a multiresolution approach, which significantly reduces the permutation misalignment while keeping the valuable spectral resolution intact. We carry out our experiments under varying acoustic conditions, and for all methods, we compare the performance to an ideal benchmark.
展开▼
