Log-spectral feature reconstruction based on an occlusion model for noise robust speech recognition.