Sopro
samuel-vitorino/soproReviewed by AI
Checked by RepoPicks Radar on
What is Sopro?
Sopro is a small text-to-speech model (120 million parameters) that speaks English, European Portuguese, French and German and copies a voice from 5 to 20 seconds of reference audio. The fastest way to try it is the in-browser demo page, where nothing is installed. On your own machine, `uvx --from sopro soprotts serve` opens a local page at http://localhost:7860; it also works from the command line or from Python. It runs on a laptop processor, uses a graphics card when there is one, and downloads the model on first use.
Sopro is free and open source (Apache-2.0) and can replace ElevenLabs, Murf, and Fish Audio.
How easy is it to install Sopro?
A · download and open · Web
- Windows C
- macOS C
- Linux C
- Web A
The browser demo (samuel-vitorino.github.io/sopro) is an A: you open the page and the model loads by itself, though the README warns that on phones the quantized version is weaker and low-memory devices "may crash". On Windows, macOS and Linux it is a C: install Python or uv, run `uvx --from sopro soprotts serve` or `pip install -U sopro`, and wait for the model download.
Sopro on GitHub
- Stars
- 983on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +5stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux, Web
- You need to install
- Python
- To build from source
- Python, Graphics card
- Disk space
- 900 MB–4.2 GB · estimate
- Accounts or keys
- None
- Cost
- Free · Free software under Apache-2.0 with free model files; the cost is your own machine plus roughly 0.9 to 4.3 GB of disk for the downloaded model.
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, Murf, Fish Audio
- Last change on GitHub
- Sep 1, 2026
Sopro risk signals (5)
- Attentioncopies a voice and no watermark
Main features: "Zero-shot voice cloning from 5-20 seconds of reference audio". Disclaimers: "We did not add watermarking: with an open-source inference pipeline it would be trivial to remove... Please use the model for good: do not impersonate people." Generated audio therefore carries no mark showing it was made by a machine.
AI review
- Infomodel downloads on first use
"The model downloads on first use and stays in the Hugging Face cache", from huggingface.co/samuel-vitorino/sopro-v2-turbo; the repository itself ships no model file.
AI review
- Infono github releases only package registries
The project facts say "GitHub releases: 0"; versions are published only as `pip install -U sopro` (pyproject.toml declares version 2.2.0) and `npm install @soprotts/onnx-web@latest`, with no checksum in the repository.
AI review
- Infoauthor documents quality limits
Disclaimers say the text frontend is "deliberately minimal", so "some abbreviations, numbers, and symbols may not be pronounced correctly" (write `one plus two`), mixed-language sentences "can be mispronounced", and the streaming path is "not bit-exact with the offline path".
AI review
- Infotraining code not released
Disclaimers: "We are not planning to release the training code in the near future due to its complexity." The repository holds only the code that runs the model, and the training data listed is Emilia YODAS, LibriTTS-R and FalAR.
AI review
Evidence is shown in its original language.
What Radar could not check about Sopro
- how close the cloned voice sounds to the original recording
- whether the speed figures (about 300 ms to first audio on a laptop processor) hold on your machine
- how much disk the model really takes after the first download
- whether the browser demo works on your phone or runs out of memory
- how well the three languages other than English are pronounced