Comparing Performances of Different Multiband Adaptive Architectures for Noise Reduction

The principal aim of this work is to test the performances of two different adaptive sub-band binaural structures in the active noise reduction problem. The first structure is realized by a weighted overlap and add (WOLA) DFT bank, with fast recursive least square adaptation algorithm. The second structure is based on a particular critically sampled filter bank with NLMS algorithm; using a proper choice for the analysis filters, it is possible to consider extra terms in order to avoid aliasing distortion. The multi microphone sub-band adaptive (MMSBA) scheme is used as a term of comparison representing one of the most efficient sub-band scheme known in literature for noise reduction, both from computational complexity and from performance evaluation point of view. The use of Wiener filtering for the reduction of residual incoherent noise is considered as a subsequent step. Performance evaluations are made, using the same computational complexity for each structure, applying two different measure to determine both noise reduction and enhanced speech distortion