TTS Spaces Arena

TTS Spaces Arena

The codebase for TTS Arena v2, a Gradio app where you vote blind on speech from text-to-speech models on Hugging Face. For people who want to compare voices by ear or run their own arena.

Made by Pendrokar on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: TTS Spaces Arena is the codebase for TTS Arena v2, a Gradio app where people vote blind on the speech that text-to-speech models on Hugging Face produce. The Space names models such as MaskGCT, MegaTTS3, XTTS-v2 and two versions of fish-speech. Two companion datasets hold the saved votes and a tracker of TTS models.

Who it suits: It suits people who would rather judge voices by ear than by a benchmark number.

Title: Who might use this arena
Industry | Who uses it | How they use it | Example
Speech research | Researchers | Compare text-to-speech models by ear | Vote on which unnamed clip sounds better
Voice products | Developers | Shortlist a TTS model for a feature | Listen to several models read the same sentence
Open-source speech | Maintainers | Run an arena of their own | Start the code locally and try the voting flow

If you need to generate speech for a product, you can skip it, since this is built for comparing and voting.

Where to use it: You can use it online or run a copy yourself.

• Web browser: the Space's page on Hugging Face, with nothing to install
• Windows, macOS or Linux: a local copy that runs with Python
• Hugging Face Spaces: where the hosted arena runs, on the cpu-basic hardware tier

Title: What the arena is good for
Use case | Who | What they get back
Judge voices blind | Anyone curious | A fair listen, with the model names hidden
Compare TTS models | Researchers | Votes that record which speech people preferred
Run your own arena | Maintainers | A local copy of the TTS Arena v2 code

How to set it up: The hosted arena needs nothing. A local copy needs Python and pip.

• Python, with pip
• The system packages that packages.txt lists, installed with apt-get
• A Hugging Face token set as HF_TOKEN is optional, because anonymous accounts may face more limits

1. Install the huggingface_hub package.
2. Download the Space's files into a folder.
3. Move into that folder.
4. Install Gradio 6.5.1.
5. Install the packages from requirements.txt.
6. Start the app with Python.

pip install -U huggingface_hub
hf download Pendrokar/TTS-Spaces-Arena --repo-type space --local-dir TTS-Spaces-Arena
cd TTS-Spaces-Arena
pip install gradio==6.5.1
pip install -r requirements.txt
python app.py

How to use it: Open the Space online and listen to the speech the models produce, then cast your vote while their names stay hidden. For a local run, set RUNNING_LOCALLY to 1, which stops the app from syncing with the database.

RUNNING_LOCALLY=1 python app.py

Title: Parts of the repository
Tool | What it does
app.py | Starts the Gradio arena
app | A folder of components, with its own notes for more detail
voice_samples | A folder of voice sample files
harvard_sentences.txt | A list of test sentences, going by its name
play_tts_space.py | A script whose name points to playing speech from a TTS Space

Example: You open the arena and listen to one sentence spoken by two unnamed models. You vote for the one that sounds better, and your vote is recorded. A local run with RUNNING_LOCALLY set skips the database sync.

For newcomers and experts: If you are new to this, vote on a few rounds and trust your ears. Names stay hidden, so you judge the sound and not the brand.

If you know the field, the app is a Gradio 6.5.1 project with several components under the app folder. The repository also carries sentence lists, among them Harvard sentences and expression sentences, and text files tied to Llama 3 and Command R prompting.

Open the arena
Listen to speech from hidden models
Cast a blind vote
The vote is saved to a dataset

Good to know: A local run needs one environment variable set.

• Local runs: RUNNING_LOCALLY must be 1, or the app syncs with the database
• Token: HF_TOKEN may be the only other variable you need
• Hosting: the hosted copy runs on cpu-basic hardware
• License: zlib
• Cost to run: none named, and a Hugging Face token is optional
• Data it reaches: the speech models it names, the votes it records, and the Python and system packages it installs

On Hugging Face, 13 November 2026

• Likes: 493
• Built with: gradio
• License: zlib

Links

• Space: https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena
• Files: https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.