Overview

Keyword spotting (KWS) lets your device recognize and respond to specific spoken commands. The Nordic Edge AI Lab pipeline automatically trains a keyword spotting model and builds ready-to-use firmware for the nRF54LM20B SoC with Axon NPU. You only need to enter your commands — the rest is handled automatically.

A command is a short phrase your device responds to when spoken, such as "Play". Unlike wake word detection, which listens for a single trigger phrase, keyword spotting recognizes a set of commands and reports which one was detected.

The keyword spotting pipeline requires no data collection. You provide the commands as text, and the Edge AI Lab trains a classification model to recognize each command in audio and produces a model archive that runs on the Axon NPU. This removes the need to record and label thousands of audio samples and lets you evaluate voice commands early in development.

The workflow consists of three steps:

1. Enter commands — Add the commands your device must recognize.
2. Start training — Training starts automatically and typically takes 2–3 hours, depending on the number of commands. In some cases it may take longer.
3. Run on device — Download the ready-to-use firmware package and run it on your device.

You can safely leave the page after starting training. The results become available once training is complete.