Not All Weights Are Created Equal: Enhancing Energy Efficiency in On-Device Streaming Speech Recognition