Skip to content

Model Export for Keyword Spotting (KWS)

The export step in soundKIT converts your trained KWS model into formats optimized for embedded deployment, including TensorFlow Lite (TFLite) and C header files for use with Ambiq SDKs.

Running Export

To export the best-performing model checkpoint:

soundkit -t kws -m export -c configs/kws/kws.yaml

Configuration (export section of kws.yaml)

export:
  epoch_loaded: best
  tflite_dir: ${job_dir}/tflite

Output Formats

  • TFLite (.tflite): For use in TensorFlow Lite runtimes or mobile applications.
  • C Header (.h): For direct integration into embedded C applications, particularly on Ambiq EVB platforms.

Exported files will be saved under the specified tflite_dir.

Notes

  • Make sure your model has been trained and saved before running this step.
  • The epoch_loaded should match the best or final checkpoint you intend to deploy.
  • The exported files can be directly used with the EVB or PC demos.

Refer to the Demo Guide to run real-time keyword detection using the exported model.