Using missing feature theory to actively select features for robust speech recognition with interruptions, filtering and noise KN-37

Speech recognizers trained with quiet wide-band speech degrade dramatically with high-pass, low-pass, and notch filtering, with noise, and with interruptions of the speech input. A new and simple approach to compensate for these degradations is presented which uses mel-filter-bank (MFB) magnitudes as input features and missing feature theory to dynamically modify the probability computations performed in Hidden Markov Model recognizers. When the identity of features missing due to filtering or masking is provided, recognition accuracy on a large talker-independent digit recognition task often rises from below 50% to above 95%. These promising results suggest future work to continuously estimate SNR's within MFB bands for dynamic adaptation of speech recognizers.