How do you handle privacy for voice sessions?

Voice-first makes so much sense for moments when typing feels like too much. Curious how you handle voice data: is audio stored, or only transcripts? And are sessions used for model training? For many people that's the deciding factor before they open up about something personal.