On Over-Determined Frequency Domain BSS

This paper introduces two new frequency domain overdetermined blind source separation (BSS) algorithms: Inter-frequency Correlation with Microphone Diversity (ICMD), and ICA with Triggered Principal component analysis (ITP). In the first, we consider different sets of microphones, where in each set the number of microphones and sources are equal. In the second, we extract principal components from an overdetermined mixture to form a determined mixture for separation. Both techniques utilize inter-frequency correlation to align permutations via energy profiles. Both monitor the condition number of an inter-frequency cross-correlation matrix of the normalized de-mixed signals´ envelopes to determine if separation has failed for the current ICA input configuration; if so, the input configuration is revised and efficiently realigned to produce a better mixture for separation. The complexities and performances of these algorithms are examined in both simulations and a real-room measurement, with three and five sources. They are also compared to other recent frequency domain BSS algorithms for benchmarking purposes. Results show that generally, ICMD and ITP show similar performance with each other and with one of the benchmarking algorithms. However, ICMD is more computationally efficient.