Binaural Bark subband preprocessing of nonstationary signals for noise robust speech feature extraction

A two channel approach to noise robust feature extraction for speech recognition in the car is proposed. The coherence function within the Bark subbands of the mel-frequency-cepstral-transform is calculated to estimate the spectral similarity of two statistic processes. It is illustrated how the coherence of speech in binaural signals is used to increase the robustness against incoherent noise. The introduced preprocessing method of nonstationary signals in two microphones results in an additive correction term of the mel-frequency-cepstral-coefficients.