Enhanced IVA for audio separation in highly reverberant environments
thesisposted on 15.04.2019, 15:44 by Suleiman Erateb
Blind Audio Source Separation (BASS), inspired by the "cocktail-party problem", has been a leading research application for blind source separation (BSS). This thesis concerns the enhancement of frequency domain convolutive blind source separation (FDCBSS) techniques for audio separation in highly reverberant room environments. Independent component analysis (ICA) is a higher order statistics (HOS) approach commonly used in the BSS framework. When applied to audio FDCBSS, ICA based methods suffer from the permutation problem across the frequency bins of each source. Independent vector analysis (IVA) is an FD-BSS algorithm that theoretically solves the permutation problem by using a multivariate source prior, where the sources are considered to be random vectors. The algorithm allows independence between multivariate source signals, and retains dependency between the source signals within each source vector. The source prior adopted to model the nonlinear dependency structure within the source vectors is crucial to the separation performance of the IVA algorithm. The focus of this thesis is on improving the separation performance of the IVA algorithm in the application of BASS. An alternative multivariate Student's t distribution is proposed as the source prior for the batch IVA algorithm. A Student's t probability density function can better model certain frequency domain speech signals due to its tail dependency property. Then, the nonlinear score function, for the IVA, is derived from the proposed source prior. A novel energy driven mixed super Gaussian and Student's t source prior is proposed for the IVA and FastIVA algorithms. The Student's t distribution, in the mixed source prior, can model the high amplitude data points whereas the super Gaussian distribution can model the lower amplitude information in the speech signals. The ratio of both distributions can be adjusted according to the energy of the observed mixtures to adapt for different types of speech signals. A particular multivariate generalized Gaussian distribution is adopted as the source prior for the online IVA algorithm. The nonlinear score function derived from this proposed source prior contains fourth order relationships between different frequency bins, which provides a more informative and stronger dependency structure and thereby improves the separation performance. An adaptive learning scheme is developed to improve the performance of the online IVA algorithm. The scheme adjusts the learning rate as a function of proximity to the target solutions. The scheme is also accompanied with a novel switched source prior technique taking the best performance properties of the super Gaussian source prior and the generalized Gaussian source prior as the algorithm converges. The methods and techniques, proposed in this thesis, are evaluated with real speech source signals in different simulated and real reverberant acoustic environments. A variety of measures are used within the evaluation criteria of the various algorithms. The experimental results demonstrate improved performance of the proposed methods and their robustness in a wide range of situations.
Tripoli University (Libya)
- Mechanical, Electrical and Manufacturing Engineering