Attribute-based histogram equalization (HEQ) and its adaptation for robust speech recognition

Histogram equalization (HEQ) is a simple and effective feature normalization technique for robust speech recognition. Recently, we proposed to adapt HEQ transform to each test utterance using a maximum likelihood (ML) criterion and observed improved performance. In this paper, we further the study by applying attribute-based HEQ and its ML adaptation. Instead of applying a global HEQ transform to the test utterance, we propose to apply different HEQ transforms to the 6 manners of speech, e.g. vowel and fricative. We also developed the ML adaptation algorithm of the attribute-based HEQ. Experimental results show that the attribute-based HEQ adaptation obtained 21.8% and 19.5% relative error rate reduction over the global HEQ baseline on the Aurora-2 and Aurora-4 benchmarking tasks, respectively.