Dimensional reduction using Blind source separation for identifying sources

Separation of independent sources using Blind Source Separation (BSS) techniques requires prior knowledge of the number of independent sources. Performing BSS when the number of recordings is greater than the number of sources can give erroneous results. Techniques employed to estimate suitable recordings from all the recordings require estimation of number of sources or require repeated iterations. This paper demon-\nstrates that normalised determinant of the global matrix is a measure of the number of independent sources, K, in a mixture of M recordings. This paper also shows that performing ICA on K out of M randomly selected recordings gives good quality of separation. The qualities of the outcome of this experiment were measured using Signal to Interference Ratio (SIR) and Signal to Noise Ratio (SNR). The results demonstrate that using\nthis technique, there is an improvement in the quality of separation as measured using SIR and SNRs.