memo · ADR 0013 · supersedes 0006

13.Hosted speech to text, after planning to self-host.!?

The problem: Voice notes have to become text, well, in the background, for one user who talks to the app every day.
The options: Whisper on a homelab; Whisper on a rented GPU; a hosted API billed by the minute.
The move: Hosted. The homelab the first decision assumed never existed, and the privacy argument for self-hosting went with it. A provider with good timestamps and a simple contract won.
What we accepted: Audio leaves the box, and the cost scales with how much I talk. The earlier decision was superseded on purpose, in writing, rather than left standing while the code did something else.
When we would revisit: If the app gains other users and the per-minute bill becomes the biggest line, or if a provider's quality drifts on the eval set.