TTS: custom voices & takedowns (operators)

Operator runbook, not a viewer guide

This page is for Pyre operators moderating user-created custom voices. It is not linked from the Pyre.Bot guides sidebar. The streamer-facing side of the feature lives in Text to speech → Your custom voice.

Any user can create one custom voice, trained from a clip of their own voice through Fish Audio. A custom voice is a private Fish model (it never enters Fish’s public library), owned by exactly one user, and it resolves only on that user’s own channel. It is provider-hidden everywhere a streamer or viewer can see it (they see the voice’s name, never “Fish”), but the admin surface shows the provider and the model id, because moderation needs them.

What the user agreed to

Creation is blocked until the user ticks a rights confirmation, and that consent is recorded before the Fish model is trained, so every custom voice carries an auditable record of which disclaimer version the user accepted and when. The v1 disclaimer states that the voice must be the user’s own or one they have explicit permission to clone, that they will not clone anyone else without consent, and that Pyre may remove the voice at any time, including on a rights or DMCA complaint, without notice.

The Custom Voices admin list

The TTS admin console has a Custom Voices card listing every custom voice on the platform: name, owner, chat tag, status, when it was created, a ▶ to play the stored reference sample, and (while the voice is still live) a Take down button.

A custom voice is named after its owner’s handle and always carries the chat tag streamer, so the name column reads as the owner and every row shares one tag. That word is reserved platform-wide: the voice whitelist refuses it as a prefix or an alias, because streamer: must always mean “the voice of whoever owns this channel”.

Status is one of:

• Active: trained and usable on the owner’s channel.
• Training: the Fish model is being trained (usually seconds).
• Failed: training did not complete; the owner can retry.
• Taken down: an operator removed it (see below). It no longer resolves.

Taking a voice down

1. Find the voice in the Custom Voices list (the ▶ sample and the owner are there to help you confirm the complaint).

2. Click Take down and confirm.

3. Pyre deletes the Fish model and flips the voice to Taken down.

Takedown is reactive by design: voices go live the moment they finish training, with no pre-publish review, and this list is how a rights or DMCA complaint is actioned. A taken-down voice stops resolving immediately: a viewer who asks for it by tag is refused, and on a channel-point or points redeem is refunded, the same path any unavailable voice takes. The charge boundary is never crossed for a voice that cannot speak.

Takedown is robust to a Fish error

The takedown flips the voice’s status even if the Fish delete call fails. The point is to stop the voice from resolving, which the status change alone guarantees. A Fish-side failure is logged, not fatal; re-running the takedown is safe (deleting an already-gone model is treated as success).

The row is kept as evidence

A takedown does not delete the record. The row stays with the recorded consent version, the acceptance timestamp, the reference sample, and which operator took it down and when: the audit trail a rights dispute may need.