Enhanced independent vector analysis for speech separation in room environments

Abstract

The human brain has the ability to focus on a desired sound source in the presence of several active sound sources. The machine based method lags behind in mimicking this particular skill of human beings. In the domain of digital signal processing this problem is termed as the cocktail party problem. This thesis thus aims to further the eld of acoustic source separation in the frequency domain based on exploiting source independence. The main challenge in such frequency domain algorithms is the permutation problem. Independent vector analysis (IVA) is a frequency domain blind source separation algorithm which can theoretically obviate the permutation problem by preserving the dependency structure within each source vector whilst eliminating the dependency between the frequency bins of di erent source vectors. This thesis in particular focuses on improving the separation performance of IVA algorithms which are used for frequency domain acoustic source separation in real room environments. The source prior is crucial to the separation performance of the IVA algorithm as it is used to model the nonlinear dependency structure within the source vectors. An alternative multivariate Student's t distribution source prior is proposed for the IVA algorithm as it is known to be well suited for modelling certain speech signals due to its heavy tail nature. Therefore the nonlinear score function that is derived from the proposed Student's t source prior can better model the dependency structure within the frequency bins and thereby enhance the separation performance and the convergence speed of the IVA and the Fast version of the IVA (FastIVA) algorithms. 4 5 A novel energy driven mixed Student's t and the original super Gaussian source prior is also proposed for the IVA algorithms. As speech signals can be composed of many high and low amplitude data points, therefore the Student's t distribution in the mixed source prior can account for the high amplitude data points whereas the original su- per Gaussian distribution can cater for the other information in the speech signals. Furthermore, the weight of both distributions in the mixed source prior can be ad- justed according to the energy of the observed mixtures. Therefore the mixed source prior adapts the measured signals and further enhances the performance of the IVA algorithm. A common approach within the IVA algorithm is to model di erent speech sources with an identical source prior, however this does not account for the unique characteristics of each speech signal. Therefore dependency modelling for di erent speech sources can be improved by modelling di erent speech sources with di erent source priors. Hence, the Student's t mixture model (SMM) is introduced as a source prior for the IVA algorithm. This new source prior can adapt according to the nature of di erent speech signals and the parameters for the proposed SMM source prior are estimated by deriving an e cient expectation maximization (EM) algorithm. As a result of this study, a novel EM framework for the IVA algorithm with the SMM as a source prior is proposed which is capable of separating the sources in an e cient manner. The proposed algorithms are tested in various realistic reverberant room environments with real speech signals. All the experiments and evaluation demonstrate the robustness and enhanced separation performance of the proposed algorithms.