Towards underdetermined source reconstruction from a clap-and-play binaural live recording

The goal of our current research is to be able to separate a few audio sources from the signals of two microphones, using a separate recording of each player clapping their hands. The separation is performed in the frequency domain, where speech and music signals are mostly sparse. Being underdetermined, the separation is performed in two steps. In the first step, the clapping is used to estimate the transfer function from each source to each microphone. In the second step, the sources are reconstructed using Second Order Cone Programming (SOCP). Our experiments show moderatly good results for synthetic mixtures (11.5dB average SNR) and poor results for the real case (2.2dB). This paper points out some of the issues that make this task a difficult one, and shows some experimental analysis of why this is so.