Privacy

TTS Arena is a volunteer-run, open-source research project. We collect what we need to keep the votes honest and to build a useful benchmark, and not much else. This page says what that is in plain terms.

What we collect

When you sign in with Hugging Face, we receive your Hugging Face user ID, username, avatar, account creation date, and the email address on the account. The account age is used for the 30-day voting requirement. The email is stored so we can contact you about your account (for example, if it's been quarantined) and is not used for anything else.

When you use the arena, we record:

• the prompts you submit and the audio each model produced for them;
• your votes, and how long you listened before voting;
• your IP address, browser user-agent, and a device fingerprint, on sign-in and with each vote;
• session events - battles started, whether they were voted on or abandoned, generation timings and errors.

A captcha may run once per browser. It stores a short-lived token and nothing about you.

Why

Almost all of this exists for one reason: a blind preference benchmark is only worth something if the votes are real. IP, fingerprint, timing, and session data are what let us spot automated voting, coordinated rings, and accounts that can somehow recognise a model they shouldn't be able to. The voting page describes how that system works.

The prompts and audio are the benchmark itself. They're what we and others can use to study how people judge synthetic speech.

What we publish

• The leaderboard: per-model ratings, uncertainty, win rates, and vote counts. Never per-voter data.
• Aggregate statistics and, from time to time, anonymised datasets of prompts, audio, and preferences. These contain no account identifiers, IP addresses, fingerprints, or email addresses.
• Accounts of integrity findings against model providers. These name the model and the provider, and can quote what the provider told us. They can describe the behaviour of accounts involved. See the Provider Agreement.

What we don't do

• We don't sell data, run ads, or share personal data with model providers. Providers see their model's ratings, not who voted.
• We don't send your identity to the model endpoints. They receive the prompt text, nothing else.
• We don't track you across other sites.

Retention

We keep vote, session, and anti-abuse records indefinitely, because a decision about the leaderboard has to remain checkable after the fact. Cached audio is kept as long as it's useful to the benchmark.

If you want your account data removed, contact us. We'll delete your profile information (username, email, avatar). Vote records may be retained in anonymised form so the leaderboard's history remains consistent.

Where it lives

The app runs on Hugging Face Spaces; the database runs on a server we operate. Hugging Face handles the sign-in and has its own privacy policy. TTS Arena is not affiliated with Hugging Face.

Contact

Questions or deletion requests: Discord or email.

Last updated 5 September 2026.