Signal processing techniques for blind source separation

The problem to be addressed is well known in the field of signal processing and to date has been tackled through the development of algorithmic techniques capable of isolating a number of different signals combined linearly in a mixture of signals. There is already a significant body of work in this field with some fairly well developed techniques, based on, for example, non-negative matrix factorisation and independent component analysis. However, a current MSc project has revealed a significant number of issues that have yet to be satisfactorily resolved. These include the poor performance of optimisation algorithms that become trapped in local minima and the unsatisfactory reconstruction of recovered signal phase. Furthermore, there have been developments more recently in approaches based on the application of deep neural networks to increase the performance of source separation. Results presented to date show significant potential when applied to relatively simple tasks such as the separation of a singing voice or a single instrument. Less literature exists for multichannel separation, since the problem becomes more complex and additional spatial information has to be taken into account. Whilst this project will have plenty of potential for comparing, developing, and improving existing approaches to the more general problem, the specific focus of the work will be will on the development of techniques that can combine source separation with virtual acoustic presentation. The work will thus be highly complementary to the work undertaken to date on the EPSRC "S3A" programme grant through which this studentship is funded.