Humans can listen to others in noisy environments, but it is difficult for humanoid robot to separate the speech of individual persons from a mixture of these sounds. Known as the cocktail party problem, this challenge has intrigued scientists and engineers for more than half a century. Using the sparse characteristic of speech signals, this thesis improves three binaural audio separation algorithms in different scenarios: 1) weak target signal extraction, 2) fast separation, and 3) under-determined reverberant speech separation (i.e., binaural speech separation problem with more than two mixed sources in the presence of echoes).

First, the thesis improves a previously reported audio cancellation kernel to separate weak target signals. Our new version of the cancellation kernel achieves...[ Read more ]

Humans can listen to others in noisy environments, but it is difficult for humanoid robot to separate the speech of individual persons from a mixture of these sounds. Known as the cocktail party problem, this challenge has intrigued scientists and engineers for more than half a century. Using the sparse characteristic of speech signals, this thesis improves three binaural audio separation algorithms in different scenarios: 1) weak target signal extraction, 2) fast separation, and 3) under-determined reverberant speech separation (i.e., binaural speech separation problem with more than two mixed sources in the presence of echoes).

First, the thesis improves a previously reported audio cancellation kernel to separate weak target signals. Our new version of the cancellation kernel achieves comparable or even better results with 3000 times the speed thanks to our analytical solutions. This solution originated from our realization that the whole extraction process can be performed in the time-frequency domain by the Short-time Fourier transform.

Second, the thesis improves the degenerate unmixing estimation technique (DUET), one of the fastest algorithms in speech separation. As a binary masking technique, DUET cannot completely separate speech signals, resulting in poor performance. We applied post-filtering with multiple linear spatial filters to improve the mask separation results and successfully resulted in significantly better separation performance.

Third, the thesis improves the l₁ minimization commonly used in audio separation algorithms. Speech separation can be converted to an l₁ minimization problem that aims to minimize the l₁ norm of the reconstructed signal. We derived and test a new weighted l₁ norm and showed that it can outperform the unweighted l₁ norm. The new algorithm can be solved using the same l₁ minimization solver but converges faster than the unweighted l₁ minimization. The improved l₁ minimization algorithms have been shown to work in the presence of reverberation and with more than two mixed speech sources.