Instead of doing transcription in features folders, perhaps create a standalone API for transcription?