Robust speech recognition using compression of Mel sub-band energies and temporal filtering

The Mel-frequency cepstral coefficients (MFCC) are commonly used in speech recognition systems. But, they are highly sensitive to presence of external noise. In this paper, we propose a two-step method to compensate noise effects on MFCC. In the first step, we propose a sub-band SNR-dependent compression function for Mel sub-band energies to give higher weights to sub-bands less contaminated with noise and give lower weights to sub-bands more contaminated with noise. In the second step, we apply temporal filters to the weighted MFCCs in order to improve their temporal characteristics. Our results on Aurora2 databases show that the proposed method has higher performance than both of conventional temporal filtering methods and weighted MFCC.