XTTS v2
The cloning model everyone actually ran in 2024.
08 · 2023–
A few seconds of a person you have never trained on, then their voice on new text. The capability leaked out of VALL-E and became a product checkbox.
Playable examples
IndexTTS, XTTS, CosyVoice 2, FireRedTTS, F5-TTS, and Fish Speech all read ‘Old will is a fine fellow…’ from the same prompt speaker. That table is the clone comparison.
Zero-shot here means: no per-speaker fine-tune. A reference clip conditions the model (as a prompt, a speaker embedding, or a style vector) and the next sentence comes out in that timbre. VALL-E framed it as in-context learning. XTTS v2 made it a Coqui checkpoint you could run. OpenVoice split timbre from style so you could clone a voice and still steer accent.
IndexTTS added duration control — necessary if the clone has to fit a picture cut. Chatterbox (Resemble, MIT license, 2025–26) is the cloning model you can actually ship in a commercial product without a research-only asterisk. CosyVoice 2 does it while streaming. Spark-TTS and Qwen3-TTS keep the capability in open codec-LM form.
The ethics are the product. Every honest demo page shows a prompt clip beside the generation. This site never pretends a vendor marketing file is a clone of our bench narrator. If a card has no clip, that is not a missing feature — it is a missing right to redistribute.
IndexTTS, XTTS, CosyVoice 2, FireRedTTS, F5-TTS, and Fish Speech all read ‘Old will is a fine fellow…’ from the same prompt speaker. That table is the clone comparison.
The cloning model everyone actually ran in 2024.
Making cloned speech fit a picture cut.
Open cloning with a license you can actually ship.
Open streaming clone quality that is actually usable live.
Another open clone on the same fair sentence as IndexTTS’s neighbours.
Multilingual zero-shot cloning before XTTS was a product name.