Improved Forward Masking on a Generalized Logarithmic Scale for Robust Speech Recognition

We previously proposed a forward masking on a generalized logarithmic scale to eliminate convolutional noise as well as to suppress additive noise. While the generalized Dynamic Cepstrum derived from the masked spectrum has been robust to both noises, the robustness to convolutional noise slightly degrades as compared to masking on the logarithmic scale, and the optimal masking coefficient depends on SNR and the type of noises. This paper improves these issures by applying the variance normalization and by controlling the masking the level depending on the estimated SNR. The recognition tests using the Aurora2 database shows that the variance normalization improves the word accuracy even for the test set C, which includes MIRS distortion, from 79.2% for the logarithmic scale to 84.0% for the generaized logirithimic scale, and that the two level masking depending on SNR improves the word acculacy for speech babble noise of 10dB from 83.7% to 89.9%.