ElevenLabs compliance for managed voices #1078

Managed voices resell ElevenLabs speech inside a Hosted subscription. Three ElevenLabs agreements apply to that: the Grants program we joined, the OEM terms that govern a bundled service, and the Terms of Use and Prohibited Use Policy, which flow down to subscribers. This issue collects everything we owe under them, so paid voice can launch knowing where it stands.

Product intent. Dormouse is a developer product. Subscribers will pick any voice from the ElevenLabs Voice Library, and Dormouse will send them there to choose one. The supplier relationship is visible by design, not something we hide.

Sources were checked 2026-10-08. docs/specs/pricing.rationale.md → "Paid-launch requirements" keeps the dated evidence.

Attribution and public statements

• Display the "ElevenLabs Grants" logo at the bottom of dormouse.sh. The grant FAQ (announcement) says a recipient "must agree to display 'ElevenLabs Grants' logo at the bottom of their website". One third-party summary adds a link to ElevenLabs for at least 12 months; check the grant acceptance email or agreement for the exact duration and whether a link is required.
• Get a written OK for naming ElevenLabs in marketing. OEM 2.B(d) bars "any public statement regarding its business with ElevenLabs without ElevenLabs' prior written approval". Terms of Use 5(c) also says the ElevenLabs name and logos "may not be copied, imitated or used … without our prior written permission". The Hosted page already says "High-quality ElevenLabs speech synthesis" and "a natural ElevenLabs voice", and we plan to say "voices powered by ElevenLabs" and link to the Voice Library.
  • Our reading: a program that requires its logo on our site expects us to attribute it, so we intend to keep this copy. A one-line reply from the grant contact approving "voices powered by ElevenLabs" plus the Voice Library link would settle it.
  • Naming ElevenLabs in the Terms and Privacy policy is required by OEM 3.A and by disclosure law, and needs no approval.
• Keep marketing within OEM 2.C(f): promote the Services "in a manner that maintains the good name and reputation of both ElevenLabs and the Services".

Grant conditions

• Confirm managed voices are not "reselling text characters". The grant page says "You cannot exchange or resell your text characters" and that doing so terminates the grant "immediately". Managed voices are a bundled service under the OEM terms, not a sale of characters, but a subscription that includes TTS is close enough to ask the grant contact directly before launch.
• Plan for the grant ending. A third-party summary reports that a grant plan converts to the Free plan when the grant period ends or 10,000 credits remain. Voice Library voices are "not available via the API to free tier users", so the downgrade would break every subscriber's library voice at once. Under the Terms that is a material service reduction, which triggers 30 days' notice and pro-rata refunds. We need either a paid plan in place before the grant ends, or a monitored date with time to give that notice.
• Resolve the OEM end-user ambiguity. OEM 1.D admits Scale and higher customers, but 1.I defines end users by internal business operations, while Hosted sells to individuals, including for personal use. This is recorded as an accepted risk in Privacy and Terms for paid Hosted #997; ask the grant contact to confirm or waive it, or keep accepting the risk explicitly.

Flow-down to subscribers (Terms, #997)

Already in website/src/pages/Terms.tsx → "Managed voices and ElevenLabs":

• incorporation of the Terms of Use and Prohibited Use Policy restrictions (OEM 3.A(a));
• no agency or partnership (3.A(b));
• ElevenLabs as a third-party beneficiary (3.A(c));
• the data-processing grant (3.A(d));
• 18+ for managed voices (Terms of Use 1(a)), and government use only with ElevenLabs' prior written consent (OEM 2.B(e)).

What remains is implementation, tracked in #1074:

• Subscribers accept the managed-voice provisions before voice access starts, and the incorporated provider versions are archived with the accepted terms.

Voice Library selection

The Voice Library guide and Voice Library Addendum put no attribution or endorsement duty on users of shared voices. They do mean a subscriber-chosen voice can change under us:

• Handle voice removal. An owner can stop sharing a voice: with a notice period, the API's Get Voice reports disable_at_unix; without one, access ends immediately. The system-voice fallback covers the failure; decide whether Dormouse warns the subscriber ahead of time.
• Decide on credit multipliers. Some library voices carry a credit multiplier, so one voice choice can cost several times another against the workspace credit limit. Either cap allowed voices by multiplier or account for it in the fair-use limit.
• Expect moderation latency. Live Moderation checks the generated text and "may introduce extra latency" for some voices.
• Verify the permitted voice selection, per the gate in docs/specs/pricing.md → "Paid-launch requirements". The voice Worker today accepts any ^[A-Za-z0-9]{1,64}$ id (VOICE_ID in hosted/server/voice.ts).

Privacy and training

These are already gates in docs/specs/pricing.md; listed here so the ElevenLabs picture is complete:

• Recheck the workspace model-training opt-out and the applicable DPA before enabling paid voices.

Done when

Every box above is checked, or explicitly accepted as a risk with a note on this issue, and ElevenLabs’ written replies are saved with the grant records.