Matching training and test data distributions for robust speech recognition

Abstract

In this work normalization techniques in the acoustic feature space are studied that aim at reducing the mismatch between training and test by matching their distributions. Histogram normalization is the first technique explored in detail. The effect of normalization at different signal analysis stages as well as training and test data normalization are investigated. The basic normalization approach is improved by taking care of the variable silence fraction. Feature space rotation is the second technique that is introduced. It accounts for undesired variations in the acoustic signal that are correlated in the feature space dimensions. The interaction of rotation and histogram normalization is analyzed and it is shown that the recognition accuracy is significantly improved by both techniques on corpora with different complexity, acoustic conditions, and speaking styles. The word error rate is reduced from 24.6% to 21.8% on VerbMobil II, a German large vocabulary conversational speech task, and from 16.5% to 15.5% on EuTrans II, an Italian speech corpus of conversational speech over telephone. On the CarNavigation task, a German isolated-word corpus recorded partly in noisy car environments, the word error rate is reduced from 74.2% to 11.1% for heavy mismatch conditions between training and test.