On-device recognition means audio does not have to be uploaded to the cloud, theoretically reducing the risk of leakage, but if some step of the product relies on cloud translation, audio clips may still leave the local machine.