Hi there,

Nordic recently shipped the nRF54LM20B with their new Axon NPU, a hardware ML accelerator built into the SoC that runs short-vocabulary speech models 15× more energy-efficiently than the Cortex-M33 on the same chip.

So I built something to find out what "always-on voice on a wearable power budget" actually costs in mA on a real bench.

The result is a hands-free voice-controlled workout tracker running entirely on the nRF54LM20 Development Kit. Say "Okay Nordic, Go" to start a workout. "Okay Nordic, Stop" then "Yes" saves it (or "No" discards). State changes hit a phone over Bluetooth LE. No cloud, no model training.

The post walks through:

• Axon NPU vs. Neuton what each one is and when to use them

• The firmware Bluetooth LE GATT service and state machine on top of Nordic's prebuilt wake-word + KWS sample

• A measured power profile the 971 µA listening floor decomposed into Bluetooth LE advertising, PDM streaming, and Axon inference

• A ~7-minute video walkthrough of the live demo, hardware setup, and three save/discard scenarios

Where the Power Actually Goes

The Axon NPU is what keeps the always-on floor under 1 mA. Running the same wake-word algorithm on the Cortex-M33 CPU would push it above 1.5 mA, which is the 15× efficiency gap in practical terms.

The rest of the budget is the PDM mic path: ~655 µA of the 971 µA listening total, vs. under 100 µA for the Axon wake-word inference. The path to dropping further isn't a smaller model, it's breaking the always-on PDM assumption (button press, motion sensor, low-power voice-activity detector).

If you're building anything that listens continuously on a battery (wearables, smart-home controls, industrial PPE), that decomposition matters.