Continual training of models based on usage data

[dscripka]

Once a user starts utilizing openWakeWord, it would be useful to have the model capture true-positive activations and use them to continually improve performance. Some potential ideas around this concept that might be worth exploring

• On-device training with TFlite: https://www.tensorflow.org/lite/examples/on_device_training/overview
• Using the shared audio embedding layer and k-nearest neighbor classifier for true-positive clips jointly with the pre-trained model
• Implementing a lightweight speaker identification model to enable a lower score threshold for the pre-trained model (i.e., increase effective recall without re-training the model)