XTTS-v2 (Voice Cloning)
Coqui's multilingual voice cloning model, abandoned upstream, kept running by the Idiap fork; non-commercial only.
About
XTTS-v2 was Coqui's flagship release in late 2023 and, for a while, the open model people reached for to clone a voice: a few seconds of reference audio was enough for it to speak in 17 languages, including across languages, with streaming output quick enough for voice assistants. It shipped inside the coqui-ai/TTS library along with fine-tuning recipes. Coqui shut down in January 2024, the original repository stopped at a February 2024 commit, and the weights on Hugging Face have not changed since December 2023. The model still works well, and the practical way to run it now is the Idiap Research Institute fork, published on PyPI as coqui-tts, which keeps XTTS working on current Python and PyTorch, with fixes as recent as October 2026; it runs best on an NVIDIA GPU. The big caveat is licensing: the weights use the Coqui Public Model License, which allows only non-commercial use, and with the company closed there is no clear way to buy the commercial license it once sold. For new or commercial cloning, consider Chatterbox from Resemble AI (MIT) or CosyVoice (Apache-2.0).
Should you use XTTS-v2 (Voice Cloning)?
Pick it when
Pick XTTS-v2 for non-commercial work such as research, personal audiobooks, or prototypes that need a cloned voice speaking any of 17 languages from about six seconds of audio, including cross-language output, on a 4 GB GPU.
Look elsewhere when
Skip it for anything commercial: the Coqui Public Model License restricts that use, and Coqui shut down in early 2024, leaving upkeep to community forks. OpenVoice is MIT licensed with six native languages. To convert existing audio, use RVC.
Alternatives to XTTS-v2 (Voice Cloning)
- Chatterbox TTS
Permissively licensed zero-shot cloning with emotion and accent control and a low-latency Turbo variant, fit for commercial agents, though rated harder to set up.
- F5-TTS
Fast flow-matching zero-shot cloning with a Gradio UI and TensorRT runtime, but its checkpoints are CC-BY-NC too, so it does not fix the license problem.
- XTTS
Same model through the full Coqui toolkit, adding streaming under 200 ms and fine-tuning recipes, with the same commercial restriction on the weights.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Voice Cloning & Voice Conversion
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Easy (2/5)
- License
- Coqui Public Model License 1.0.0 (non-commercial)
- Minimum VRAM
- 4 GB
- Added
- Apr 3, 2026