TTS Arena: Benchmarking TTS Models on HuggingFace
- Press ⚡ to quickly get cached sample pairs you've yet to vote on. (Fast 🐇)
- Or press 🎲 pick a random sentence from a prepared list. (Slow 🐢)
- Or input text (🇺🇸 English only) to synthesize audio. (Slowest 🐌)
- Listen to the two audio clips, one after the other and then vote on which audio sounds more natural to you.
- Model names are revealed after the vote is cast.
- You can use a [hotkey] for quicker voting.
Vote to help the community determine the best text-to-speech (TTS) models.
The leaderboard displays models in descending order of how natural they sound (based on votes cast by the community).
Important: In order to help keep results fair, the leaderboard hides results by default until the number of votes passes a threshold. Tick the Reveal preliminary results to show models without sufficient votes. Please note that preliminary results may be inaccurate. This dataset is public and only saves the hardcoded sentences while keeping the voters anonymous.
Show all models, including models with very few human ratings.
The TTS Arena evaluates leading speech synthesis models. It is inspired by LMsys's Chatbot Arena.
Motivation
The field of speech synthesis has long lacked an accurate method to measure the quality of different models. Objective metrics like WER (word error rate) are unreliable measures of model quality, and subjective measures such as MOS (mean opinion score) are typically small-scale experiments conducted with few listeners. As a result, these measurements are generally not useful for comparing two models of roughly similar quality. To address these drawbacks, we are inviting the community to rank models in an easy-to-use interface, and opening it up to the public in order to make both the opportunity to rank models, as well as the results, more easily accessible to everyone.
The Arena
The leaderboard allows a user to enter text, which will be synthesized by two models. Only after fully listening to each sample, the user can vote on which model sounds more natural. Due to the risks of human bias and abuse, model names are revealed only after a vote is submitted.
Credits
Thank you to the following individuals who helped make this project possible:
- VB (Twitter / Hugging Face)
- Clémentine Fourrier (Twitter / Hugging Face)
- Lucain Pouget (Twitter / Hugging Face)
- Yoach Lacombe (Twitter / Hugging Face)
- Main Horse (Twitter / Hugging Face)
- Sanchit Gandhi (Twitter / Hugging Face)
- Apolinário Passos (Twitter / Hugging Face)
- Pedro Cuenca (Twitter / Hugging Face)
* You are currently in a cloned/forked HF space of TTS-AGI/TTS-Arena
Request a model
Please create a Discussion to request a model.
Privacy statement
We may store text you enter and generated audio. We store a unique ID for each session. You agree that we may collect, share, and/or publish any data you input for research and/or commercial purposes.
License
Generated audio clips cannot be redistributed and may be used for personal, non-commercial use only. Random sentences are sourced from a filtered subset of the Harvard Sentences and also from KingNish's generated LLM sentences.
🔐 Closed Source/Weights TTS
- Microsoft Edge TTS
- MARS 6
🔓 Open Source/Weights TTS - capabilities table
See the full dataset itself for the legend and more in depth information of each model.
If you use this data in your publication, please cite us!
Copy the BibTeX citation to cite this source:
@misc{tts-arena,
title = {Text to Speech Arena},
author = {mrfakename and Srivastav, Vaibhav and Fourrier, Clémentine and Pouget, Lucain and Lacombe, Yoach and main and Gandhi, Sanchit},
year = 2024,
publisher = {Hugging Face},
howpublished = "\url{https://huggingface.co/spaces/TTS-AGI/TTS-Arena}"
}
Please note that all generated audio clips should be assumed unsuitable for redistribution or commercial use.