Retrain RNN classifier weights to include reverberated speech