Welcome to TTS Arena Hebrew

TTS Arena Hebrew is a community-driven platform for evaluating text-to-speech models on Hebrew. Built and maintained by the ivrit-ai team, this arena lets anyone compare how well different TTS engines handle Hebrew speech — including mixed Hebrew-English text, formal language, conversational tone, and more.

This project is based on TTS Arena V2 by TTS-AGI and the Hugging Face team. We are grateful for their work in creating the original open-source arena platform.

Why Hebrew?

Hebrew presents unique challenges for text-to-speech: right-to-left script, vowelless writing, frequent code-switching with English loanwords and brand names, and a relatively small speaker base compared to major languages. Most TTS benchmarks focus on English, leaving Hebrew underserved.

TTS Arena Hebrew fills this gap by providing a dedicated evaluation space where the Hebrew-speaking community can directly assess and rank how well different models handle real-world Hebrew text.

How It Works

The concept is simple: enter Hebrew text (or pick a random sentence from our curated dataset), and two anonymous TTS models will synthesize it. Listen to both, then vote for the one that sounds more natural. Model identities are revealed only after you vote.

  • Enter your own Hebrew text or pick from categorized sentences (support, podcast, news, conversation, formal, emotional, mixed)
  • Listen to two anonymous TTS models synthesize the same text
  • Vote for the model that sounds more natural, clear, and expressive
  • Track model rankings on the leaderboard using Elo ratings

Frequently Asked Questions

How are models ranked?
Models are ranked using an Elo rating system, similar to chess rankings. When you vote for a model, its rating increases while the other model's rating decreases. The magnitude of change depends on the current ratings of both models.
Can I use mixed Hebrew and English text?
Yes! Mixed text is common in everyday Hebrew (brand names, tech terms, etc.). The input just needs to contain at least one Hebrew character. We even have a "Mixed" sentence category specifically for this.
Can I suggest a model to be added?
Absolutely. If you know of a TTS model that supports Hebrew, reach out to the ivrit-ai team on Hugging Face. The arena uses a plugin architecture that makes adding new providers straightforward.
Do I need to log in?
Yes, a Hugging Face account is required to generate audio and vote. This helps prevent abuse and lets you track your personal voting history.
What are the sentence categories?
Sentences are organized by use-case: Support (customer service), Podcast (broadcast style), News (headlines and reports), Conversation (casual speech), Formal (business/legal), Emotional (expressive), and Mixed (Hebrew with English loanwords). You can filter by category when picking a random sentence.

Credits

Built by the ivrit-ai team.

Based on TTS Arena V2 by mrfakename, Vaibhav Srivastav, Clémentine Fourrier, Lucain Pouget, Yoach Lacombe, and the TTS-AGI / Hugging Face team.

Citation

If you use TTS Arena Hebrew in your research, please cite both the Hebrew arena and the original:

@misc{tts-arena-hebrew, title = {TTS Arena Hebrew: Benchmarking Hebrew Text-to-Speech Models}, author = {ivrit-ai}, year = 2025, publisher = {Hugging Face}, howpublished = "\url{https://huggingface.co/spaces/ivrit-ai/TTS-Arena-Hebrew}", note = {Based on TTS Arena V2 by TTS-AGI} }

Privacy

We may store text you enter and generated audio. If you are logged in, your votes are associated with your Hugging Face username. Data collected may be used for research purposes.

License

The arena code is based on TTS Arena V2, licensed under the Zlib license. Generated audio clips may not be redistributed and are for personal, non-commercial use only.