Fig. 1. Mean SNR of the clean and noisy (train station) Aurora2 (training (□), test (+)) and VIOS (training (), test (×)) data before and after the application of TDNR.

Fig. 2. Schematic overview of the feature extraction and mismatch reduction modules.

Fig. 3. Overall distribution of c₁, c₂, c₃ and c₄ derived from clean training data (solid line) and 0 dB SNR train station test data (dotted line).

Fig. 4. Kullback divergence between the distributions of the training and test data for the 12 cepstral coefficients. Distances for the baseline condition and after the application of HN. (Top) Distances after TDNR and TDNR + HN have been applied (Bottom).

Fig. 5. Overall distribution of logE derived from clean training data (solid line) and 0 dB SNR train station test data (dotted line) in the baseline condition, after the application of TDNR, after HN, and after TDNR + HN have been applied.

Fig. 6. Recognition accuracy as a function of SNR for the Aurora2 digit recognition task (top) and the VIOS CSR task (bottom). (significant increase (+), significant deterioration (−), no significant difference (empty bar) – relative to the baseline (filled bar)).

Table 1. Recognition accuracy for the Aurora2 digit recognition task after the application of TDNR

Table 2. Recognition accuracy for the Aurora2 digit recognition task after the application of HN

Table 3. Recognition accuracy for the Aurora2 digit recognition task after the application of both TDNR and HN

Table 4. Recognition accuracy after the application of TDNR

Table 5. Recognition accuracy after the application of HN

Table 6. Recognition accuracy after the application of both TDNR and HN