Data selection by sequence summarizing neural network in mismatch condition training

Data augmentation is a simple and efficient technique to improvethe robustness of a speech recognizer when deployed inmismatched training-test conditions. Our paper proposes a newapproach for selecting data with respect to similarity of acousticconditions. The similarity is computed based on a sequencesummarizing neural network which extracts vectors containingacoustic summary (e.g. noise and reverberation characteristics)of an utterance. Several configurations of this network and differentmethods of selecting data using these "summary-vectors"were explored. The results are reported on a mismatched conditionusing AMI training set with the proposed data selectionand CHiME3 test set.