Blind Source Separation in the Time-Frequency Domain Based on Multiple Hypothesis Testing

This paper considers a time-frequency (t-f)-based approach for blind separation of nonstationary signals. In particular, we propose a time-frequency "point selection" algorithm based on multiple hypothesis testing, which allows automatic selection of auto- or cross-source locations in the time-frequency plane. The selected t-f points are then used via a joint diagonalization and off-diagonalization algorithm to perform source separation. The proposed algorithm is developed assuming deterministic signals with additive white complex Gaussian noise. A performance comparison of the proposed and existing approaches is provided.