Frequency bin nonlinear LMS adaptive noise canceler and its application to co-channel speech noise reduction

The transform domain normalized least mean square (LMS) linear filtering algorithm is extended to the nonlinear (NL) case under Gaussian filter input assumption. Properties of the transform domain NL algorithms are also investigated. By choosing a proper orthogonal transform, a faster convergence speed and a significant reduction of the numbers of NL filter weights can be obtained. A promising application is proposed to cochannel speech noise reduction. Promising results of speech intelligibility enhancement are obtained over a range of SNR between 0 dB and +or-12 dB. Spectrograms and informal listening tests of the enhanced target speech provide evidence of the proposed approach.<
>