Exploiting imbalanced textual and acoustic data for training prosodically-enhanced RNNLMs | IEEE Conference Publication | IEEE Xplore