Low Precision RNNs: Quantizing RNNs Without Losing Accuracy