OpenWhispr's underlying architecture separates speech recognition and language model processing into two independent pipelines, each of which can connect to different services or local models.