Model Compensation for Features Based on Subband Spectral Centroid Histograms

Robustness is an important issue for speech recognition systems. Mismatch between training and testing conditions due to factors such as background noise can cause severe degradation of recognition performance. One direction of research has been to find features that are less dependent on the acoustic environment, while retaining good discriminative properties. One type of such features is called subband spectral centroid histograms (SSCH). Another approach to the robustness problem has been to use information about the acoustic environment to modify the acoustic models such that they are matched to the noisy speech. Such methods are called model compensation methods, and they are usually based on traditional MFCC features. In this paper we investigate model compensation for SSCH features. Experiments on the Aurora 2 database show that model compensation approaches can be used to improve the performance for SSCH features, although the gains in performance are not as great as those obtained when compensating MFCC models