Your voice stays private.

Speaker permission

Voice registration records the speaker’s name, permission type and acceptance of the voice consent terms. When uploading another speaker’s recording, keep written permission covering cloning and your intended use, and include its reference when you register.

With Invite a friend, the speaker confirms permission on the recording page. That submission is linked to the voice’s consent record; you do not need to enter a separate permission reference. Revoking an unused invitation disables its recording link. To remove a voice that has already been created, delete that voice separately.

Do not use a voice to impersonate someone without permission or mislead listeners about who is speaking.

Delete and revoke

Deleting a voice immediately blocks new speech and access to its previous downloads. The console displays “Deleting” while reference recordings, voice prompts and retained audio are removed. It shows “Deleted” after cleanup succeeds.

Private storage

Reference recordings, voice prompts and retained audio in the API’s private object store are encrypted with AES-256-GCM and accessible through your authenticated project. The inference service processes decoded audio and retains private voice-model files until voice deletion completes. Host access controls protect these working files; they do not use the API object encryption.

Operational request logs record identifiers, character counts and status, without storing the source text or audio in those logs.

Retention

Generated audio and recoverable job input expire within 24 hours. Voice references and prompts remain until you delete the voice. Payment, usage and consent records are kept separately from audio so charges and permissions can be traced.