On Real Time Q-Log-based Feature Normalization for Distant Speech Recognition

The computation of long term mean in feature normalization methods requires information on future frames, and thus makes them inapplicable for real-time implementations. Previously, q-log spectral mean normalization (q-LSMN) as feature normalization method is proposed, and it shows more effective result than conventional normalization methods. However, q-LSMN has not yet been implemented on real time. In this paper, we propose a real time implementation of q-LSMN. In this method, the mean is updated recursively based on only previous frames, hence no future frame information is needed. Experiments on Aurora-5 databases showed that while real time q-LSMN achieved slightly worse performance than non real time q-LSMN as expected, it improved the recognition accuracy up to 54.22% compared to that of non-real time conventional normalization methods such as cepstral mean normalization (CMN) and Log spectral mean normalization (LSMN).