Training the AI
- Run script NN to split large audio files into segments. See YY for insturctions for working with e.g. Xenp-Canto -files.
- System creates 10-second audio segments and spectrograms
- Annotate audio segments
- User annotates segments using web UI
- Train the AI using
- Audio segment spectrograms
- Annotation data on database
- Test the AI
Using the AI
-
User places audio files from one night into a folder (later: more nights)
-
User starts the process
-
System creates audio segments automatically
-
AI handles each segment
-
System saves predictions to a file, with links to those segments that are above thresholds. (Later: to database, like manual annotations, so that can retrain the AI)
-
System deletes audio files of below-threshold segments, to save space. (Maybe also spectrograms?) (- Systems shows migration activity index for the night)
-
User listens and views segments with predicted NFC's
- Identifies species
- Records these as observations / no-birds to the system, which creates annotations to the db.
-
System sends observations to FinBIF Notebook via API
-
Eventually AI is retrained with the new material in the db.