On subband-based blind separation for noisy speech recognition

A method for denoising noisy speech signals in the feature extraction process for robust speech recognition is proposed. The method uses independent component analysis, in which a noise signal is linearly separated from two noisy speech microphone recordings. In addition, the method is optimized by computing a modified band that sums up FFT point values in several divided ranges of one band, and computes each band energy using the summed values. Thus, the number of unmixing networks is reduced. For instantaneous mixtures of speech and noise, the method showed the same recognition performance as for the clean speech signal case. For noisy speech signals recorded in real environments, the recognition rate was considerably increased after separation and the methods was particularly effective for a very low signal to noise ratio.