Text this: Improving the Performance and Learning Stability of Parallelizable RNNs Designed for Ultra-Low Power Applications