Provider Agreement

Effective 5 September 2026.

This agreement is between the TTS Arena project ("TTS Arena", "we", "us") and any person or organisation that submits a text-to-speech model for evaluation (the "Provider", "you"). By submitting a model, or by leaving a model listed after this agreement is posted, you accept it on behalf of yourself and the organisation you represent.

It is written in plain English on purpose. Where it says "must", it means it.

1. Who we are

TTS Arena is an independent, open-source community project maintained by volunteers. It is not affiliated with, operated by, sponsored by, or endorsed by Hugging Face, Inc., or by any model provider. The arena is hosted on Hugging Face Spaces and uses Hugging Face accounts for sign-in, in the same way any third-party project may; that does not make it a Hugging Face product, and nothing in the arena's rankings, labels, or communications is a statement by Hugging Face. The source is public under the Apache 2.0 licence.

2. Submitting a model

1. Accurate information. Everything you tell us about the model - its name, provider, whether it is publicly available, which model is actually behind the endpoint - must be true, and you must tell us when it changes.
2. Identity. You must tell us who you are: the organisation, and a contact who can speak for it. This applies to codenamed entries too (see section 5).
3. Authority. The person submitting confirms they are authorised to bind the Provider to this agreement.
4. One model, one entry. We evaluate one version of a model at a time. A new version is a new submission, and we may decline to run two entries from the same provider that are substantially the same model.
5. Our discretion. We may accept, decline, delay, or remove any submission, for any reason, including capacity, cost, and the integrity of the board.

3. The endpoint

1. It serves the model you said it does, and the same model that users of your product get. No arena-specific tuning, no better model for arena traffic, no silent swaps. If you change what is behind the endpoint, tell us first; we will decide whether the entry continues or restarts.
2. Prompt handling. The endpoint must synthesise what it is given. It must not detect arena prompts, treat arena traffic differently from ordinary traffic, or cache and replay outputs for known prompts. If it refuses a prompt (for example on safety grounds), it must return an error, not degraded audio.
3. Voices. You supply a fixed pool of voices for us to rotate through. We do not do zero-shot cloning in the arena.
4. Availability. The endpoint should be reliably available. We may time out, retire, or remove a model whose endpoint is failing, slow enough to distort the listening test, or rate-limiting the arena below what evaluation needs.
5. Cost. You bear the cost of serving your own endpoint, and you must not charge the arena for it. We may cap the volume of requests we send.
6. Keys. We keep your API credentials confidential, use them only for the arena, and delete them when the model is retired. You may rotate them at any time by telling us.

4. Integrity

The integrity policy is part of this agreement. In full:

1. No voting on your own model by anyone on your side: employees, contractors, interns, agencies, investors, family, communities you run, or anyone acting at your request or with your encouragement. This includes personal accounts and any "test" or shared accounts. You must not create, lend, or control voter accounts.
2. No un-blinding. You must not use your endpoint's request logs, timestamps, telemetry, watermarks, audio artefacts, prompt matching, or any other side channel to work out which side of a battle is yours, whether or not you then vote. The arena is a blind test; defeating the blind for yourself defeats the test.
3. No purchased, incentivised, automated, or coordinated votes. No paying or rewarding people for votes, no bots or scripts, no vote drives, no asking a community to "go vote for us".
4. No selective participation. You must not use the arena as an internal evaluation harness - for example, generating battles and abandoning the ones your model isn't in. Compare models with your own tooling, or with custom prompts here, which do not affect the ratings.
5. No harvesting. You must not scrape the arena, or use it to collect other providers' audio, prompts, or outputs for training, benchmarking, or reverse-engineering.
6. Disclosure. If you become aware that anyone connected to you has done any of the above, you must tell us promptly.
7. Cooperation. If we contact you about activity that looks like a breach, you must respond honestly and within a reasonable time. Silence, or answers that later prove false, are themselves a breach.

5. Codenamed (anonymous) entries

1. You may run an unreleased model under a codename. It is anonymous to voters, not to us: we must know the provider and the model.
2. Anonymity is bounded by time or vote count, set when the entry is accepted, and ends regardless of how the model performs. You may withdraw before then; you may not reveal only because it won, or keep an entry codenamed indefinitely.
3. A model that is already publicly available runs under its real name.
4. We keep your identity confidential until the reveal as a courtesy, not a guarantee. If this agreement is breached we may disclose the provider as part of the outcome. We may also disclose it if required by law.

6. Data, audio, and publication

1. Generated audio. Audio your endpoint produces for the arena may be stored, played to voters, cached, and published, including as part of research datasets and in illustrations of results. You grant TTS Arena a perpetual, worldwide, royalty-free licence to do so, and you confirm you have the rights needed to grant it.
2. Votes and sessions. Vote records, session data, and anti-abuse signals belong to the arena. We may publish aggregate results and anonymised data. Providers do not receive raw voter data.
3. Correspondence. Communications between you and the arena about a listing, an integrity review, or a decision may be quoted or published, in whole or in part, as part of our account of that matter.
4. Outcomes. We may publish the outcome of any integrity review, including the model name, the provider name, what the data showed, and what the provider told us. By submitting, you consent to that publication.
5. Nothing is deleted. Votes, sessions, and correspondence are retained so that any decision can be checked afterwards.

7. Using arena results

1. You may cite your model's ranking, rating, or win rate, provided you state the date and link to the leaderboard. Rankings move; a claim must be true when it is made and must not be presented as permanent.
2. You must not alter, crop, or restyle leaderboard screenshots in a way that changes their meaning, omit the Preliminary or Suspended labels, or present results of custom prompts, side-by-side listening, or your own testing as arena results.
3. You must not describe TTS Arena results as an official ranking by Hugging Face or anyone else, or imply that TTS Arena endorses your model.
4. You may use the TTS Arena name to refer to the arena factually. You may not use it, or the arena's marks, in a way that suggests partnership or endorsement.

8. Labels, withdrawal, and removal

1. Preliminary marks a model with fewer votes than we need to settle its rating. It is ranked normally.
2. Retired marks a model whose provider has stopped serving it. You may retire a model at any time by telling us. It stays on the board with the rating it earned.
3. Suspended marks a model removed from evaluation for a breach of this agreement. It stays on the leaderboard, at the bottom, unranked, with the reason and date. The label has no expiry. A suspended model cannot be retracted, withdrawn, renamed, or re-submitted under another name, and retiring it does not remove the label.
4. We may remove or retire a model for technical reasons (a failing endpoint, a discontinued API) without that being a finding against the provider, and we will say so.

9. Consequences of breach

If we find that this agreement has been breached:

1. affected votes are excluded and the leaderboard is recomputed from the clean set;
2. the model is suspended as described in section 8;
3. we may publish an account of what happened as described in section 6;
4. we may decline future submissions from the same provider, or accept them on conditions;
5. we may inform other evaluation platforms of the outcome.

Whether the breach was intended is not the test. Conduct that moves the board is a breach whether or not that was the purpose.

Findings are made by the TTS Arena team on the evidence available to us. We will contact the provider before a decision where we reasonably can, and we will consider what the provider says. Decisions are final.

10. Disclaimers

1. The arena, its ratings, and its API are provided "as is", without warranty of any kind. Ratings are statistical estimates from crowd preferences; they are not a measure of fitness for any purpose.
2. We do not guarantee that a model will be listed, evaluated to any number of votes, evaluated within any time, or kept on the board.
3. We do not guarantee uptime, and we may change the arena's methodology, thresholds, design, or rules at any time. Material changes to this agreement will be posted here with a new effective date; keeping a model listed after that is acceptance.

11. Liability and indemnity

1. To the fullest extent permitted by law, TTS Arena and its maintainers are not liable to you for any indirect, incidental, consequential, or special damages, or for lost profits, revenue, or business, arising from the arena, a listing, a rating, a label, or a published outcome. Where liability cannot be excluded, it is limited to US $100.
2. You will indemnify TTS Arena and its maintainers against claims arising from your endpoint, your audio, your use of arena results, or your breach of this agreement.

12. General

1. Governing law. This agreement is governed by the laws of the State of California, USA, without regard to conflict-of-law rules. Disputes will be brought in the state or federal courts located in California, and you consent to their jurisdiction. Before bringing any claim, you agree to raise the matter with us in writing and allow thirty days for an informal resolution.
2. Relationship. Nothing here creates a partnership, agency, employment, or joint venture. Listing a model is not an endorsement of the provider by the arena, or of the arena by the provider.
3. Assignment. You may not transfer this agreement without our written consent. We may assign it to a successor maintainer of the project.
4. Severability. If any part is unenforceable, the rest stands.
5. Entire agreement. This page, the integrity policy, and the Provider API documentation are the whole agreement between you and TTS Arena about a listing. Anything said in Discord, email, or on a call does not change it unless we confirm the change in writing.
6. Contact. The TTS Arena team can be reached on Discord or by email.