Which voice sounds more human?

Leaderboard

Leaderboard0 of 0 models · ranked by human preference; domain columns rate each model on that kind of prompt (dimmed = few votes yet). Suspended models are listed last, unranked.

Include new models

Model finder

What are you building?

RequirementsHard filters: models that miss one drop out

Language

Generation time

Not enough data yetWe're timing generations now; this filter opens once models have enough samples.

Deployment

List price, per 1M characters

About

TTS Arena ranks text-to-speech models by ear. Anyone can listen to two anonymous voices, pick the better one, and the votes become an open leaderboard.

How it works

01TypeWrite any line, or pull one from the prompt set.

02ListenTwo anonymous models read it back as A and B.

03VotePick the one that sounds more human.

04RankEvery vote feeds the fit behind the leaderboard.

Questions

How are models ranked?Each vote is a head-to-head result. The board is a single Bradley–Terry fit over every recorded matchup, ranked by a conservative lower bound so a model has to be both good and certain to place highly. A model needs 100 votes to appear and is marked preliminary until 300.

Why blind comparisons?Knowing the name colors how people hear a voice. Hiding identities until after the vote keeps the judgment on the audio.

What are the rules?Sign in with a Hugging Face account at least 30 days old. Prompts can be in any language and are capped at 1,000 characters. The main board uses English prompts; other languages get their own boards once they have enough votes.

Can I get a model added?Open an issue on GitHub or reach out on Discord. Unreleased models can be evaluated anonymously before launch.

Cite

text
@misc{tts-arena-v2,
  title  = {TTS Arena 2.0: Benchmarking Text-to-Speech Models in the Wild},
  author = {mrfakename and Srivastav, Vaibhav and Fourrier, Clémentine and
            Pouget, Lucain and Lacombe, Yoach and main and Gandhi, Sanchit and
            Passos, Apolinário and Cuenca, Pedro},
  year   = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/spaces/TTS-AGI/TTS-Arena-V2}}
}