We proposed the Complex Spectrum Circle Centroid (CSCC) method for restoring the complex spectrum of target signal from multiple microphone input signals in noisy emvironments. Unlike most existing methods, this method can handle with correlated signal and noise, non-planar wave propagations, and any type of noise without locating the noise sources or training filter coefficients. The noise reduction process is non-linear to the input and independent between different frequencies.
The proposed method was evaluated in simulated noisy speech recognition experiments and shown to be significantly effective not only in the case of single noise but also in multiple noise case. Preliminary experiments in a real reverberant environment has not yet yielded better results than the delay-and-sum method.
This new method is still an on-going work to be further explored, both theoretically and experimentally.