A common question about voice models: where do the recordings live? The honest answer is: ours don't. The voice model is trained, packaged, and shipped. We never keep the audio that trained it.
Why we chose this
Storing user audio is the kind of thing you do once and regret forever. Even with redaction, even with consent, even with deletion guarantees — it sits there. We chose not to have it sit there.
How it works in practice
Voice cloning happens in three steps. First, the user submits a 60-second sample, which is processed entirely on-device to extract acoustic features. Second, those features — not the audio — are sent to our private cluster, which trains a 50 MB model in about 90 seconds. Third, the model is returned to the user, and the cluster discards the features as soon as the model is built.
What 'discarded' means here
We don't keep logs of the features. We don't write them to disk. The container that trains the model has a 5-minute TTL and is destroyed after one job. We've thought about this carefully, and we'd rather have a slower clone pipeline than a faster one that has a copy of your voice sitting on a server.
"The best way to keep a secret is to never learn it."