Real-Time Voice Cloning
CorentinJ/Real-Time-Voice-CloningReviewed by AI
Checked by RepoPicks Radar on
What is Real-Time Voice Cloning?
This is the code behind a master's thesis: you give it a few seconds of someone speaking, it builds a digital fingerprint of that voice, and then it reads text you type in a voice that resembles it. There is a small desktop window (the toolbox) and a command-line version, and the training scripts are included. It runs on Windows and Linux, needs ffmpeg plus Python set up through uv, and an NVIDIA card makes it usable. The author himself says the project has "quickly gotten old" and that paid services sound better.
Real-Time Voice Cloning is free and open source (MIT) and can replace ElevenLabs and Respeecher.
How easy is it to install Real-Time Voice Cloning?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- Linux D
There is no release to download at all. On both Windows and Linux the README asks you to install ffmpeg, then install the uv tool, then run `uv run --extra cuda demo_toolbox.py` (or `--extra cpu` without an NVIDIA card), which sets up a Python environment and downloads the models on first use. It is terminal work on both, and the project only supports Windows and 64-bit Linux, not macOS.
Coming soon: install in 1 click with the RepoPicks app.
Real-Time Voice Cloning on GitHub
- Stars
- 60,237on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +13stars in 7 days
Stars over time
Facts
- Runs on
- Windows, Linux
- You need to install
- Python, NVIDIA graphics card, FFmpeg
- To build from source
- Python, NVIDIA graphics card, FFmpeg
- Disk space
- 900 MB–4.2 GB · estimate
- Accounts or keys
- None
- Cost
- Free · MIT code, pretrained models downloaded at no charge, and everything runs on your own machine.
- License
- MIT (Permissive)
- Can replace
- ElevenLabs, Respeecher
- Last change on GitHub
- Mar 9, 2026
Real-Time Voice Cloning risk signals (6)
- AttentionInstall command runs a script from the internet
Install command pipes a remote script into a shell: curl -LsSf https://astral.sh/uv/install.sh | sh
automatic check
- Attentionbuilt to copy a real person s voice
The project's own description is "Clone a voice in 5 seconds to generate arbitrary speech in real-time", and the README explains that it builds "a digital representation of a voice from a few seconds of audio" and then speaks any text with it. As a rule, making a recording that sounds like a real person without that person's agreement is restricted by law in several countries.
AI review
- Attentionwindows install bypasses the script policy
The Windows step is `powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"`, which turns off PowerShell's script protection and runs a script downloaded at that moment. (The Linux line does the same thing with curl.) The README does offer `pip install -U uv` as an alternative.
AI review
- Infoauthor says it is out of date
README, "Heads up": "Like everything else in Deep Learning, this repo has quickly gotten old. Many SaaS apps (often paying) will give you a better audio quality than this repository will", and it points readers to another project (Chatterbox) for current quality. There are no releases, and the pinned libraries are old (torch 1.10, Pillow 8.4.0, scikit-learn 1.0.2).
AI review
- Infoneeds exactly python 3 9
pyproject.toml sets `requires-python = ">=3.9,<3.10"`, so it runs on Python 3.9 only. The uv tool fetches that version for you, but nothing newer will work.
AI review
- Infomodels downloaded at run time
README: "Pretrained models are now downloaded automatically", from https://huggingface.co/CorentinJ/SV2TTS. What actually speaks arrives from the internet after you install, and the torch wheels come from download.pytorch.org indexes named in pyproject.toml.
AI review
Evidence is shown in its original language.
What Radar could not check about Real-Time Voice Cloning
- Whether the install still completes today, given Python 3.9 and the old pinned libraries
- How much the cloned voice really sounds like the original, which only listening can tell
- Whether it is usable without an NVIDIA card, since a CPU option exists but the speed is not stated
- Real disk use after the models and any dataset are downloaded, since the figure here is an estimate
- Whether the automatic model download still works