Enhanced independent vector analysis for audio separation in a room environment

Independent vector analysis (IVA) is studied as a frequency domain blind source separation method, which can theoretically avoid the permutation problem by retaining the dependency between different frequency bins of the same source vector while removing the dependency between different source vectors. This thesis focuses upon improving the performance of independent vector analysis when it is used to solve the audio separation problem in a room environment. A specific stability problem of IVA, i.e. the block permutation problem, is identified and analyzed. Then a robust IVA method is proposed to solve this problem by exploiting the phase continuity of the unmixing matrix. Moreover, an auxiliary function based IVA algorithm with an overlapped chain type source prior is proposed as well to mitigate this problem. Then an informed IVA scheme is proposed which combines the geometric information of the sources from video to solve the problem by providing an intelligent initialization for optimal convergence. The proposed informed IVA algorithm can also achieve a faster convergence in terms of iteration numbers and better separation performance. A pitch based evaluation method is defined to judge the separation performance objectively when the information describing the mixing matrix and sources is missing. In order to improve the separation performance of IVA, an appropriate multivariate source prior is needed to better preserve the dependency structure within the source vectors. A particular multivariate generalized Gaussian distribution is adopted as the source prior. The nonlinear score function derived from this proposed source prior contains the fourth order relationships between different frequency bins, which provides a more informative and stronger dependency structure compared with the original IVA algorithm and thereby improves the separation performance. Copula theory is a central tool to model the nonlinear dependency structure. The t copula is proposed to describe the dependency structure within the frequency domain speech signals due to its tail dependency property, which means if one variable has an extreme value, other variables are expected to have extreme values. A multivariate student's t distribution constructed by using a t copula with the univariate student's t marginal distribution is proposed as the source prior. Then the IVA algorithm with the proposed source prior is derived. The proposed algorithms are tested with real speech signals in different reverberant room environments both using modelled room impulse response and real room recordings. State-of-the-art criteria are used to evaluate the separation performance, and the experimental results confirm the advantage of the proposed algorithms.