XTTS-v2 clones a voice from just a 6-second audio sample and speaks in any of 17 languages while preserving emotional tone, running at 24kHz. Pick it when you specifically need voice cloning, not just a stock voice. It's licensed under the Coqui Public Model License, which restricts commercial use; check the exact terms and get informed consent for whoever's voice you clone before using it for anything beyond testing.
Other good fits
Cross-language voice transfer, speaking in a different language than the original sample while preserving vocal characteristics.
+ Pros
Read pros
XTTS-v2 clones a voice from a short audio sample, just 6 seconds, and can then speak new text in that cloned voice across 17 supported languages (English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese, Japanese, Hungarian, Korean, Hindi), preserving emotional tone and stylistic qualities from the reference sample even across a language change. Output runs at 24kHz.
Pick it when…
You specifically need voice cloning from a short sample, not just a stock voice from a fixed list.
You need cross-language voice transfer: speaking in a different language than the sample while sounding like the same speaker.
You have explicit, informed consent from whoever's voice is being cloned.
− Cons
Read cons
A 6-second cloning requirement is remarkably low friction, which cuts both ways: it enables legitimate uses (accessibility, localization with consent) but also makes non-consensual cloning trivially easy, so treat consent verification as mandatory, not optional. The Coqui Public Model License's commercial restrictions mean you can't simply assume a generated commercial product is covered; check current license terms directly before any commercial deployment.
Pick something else when…
You need this for a commercial product and can't work within CPML's restrictions.OpenAI TTS offers commercial usage terms with a managed API, though without cloning.
You don't need cloning, just a good stock voice.Kokoro is a free, Apache-2.0 option with a fixed voice list and no consent complexity.
You need expressive delivery more than cloning specifically.F5-TTS is built around expressive synthesis; check its own license too.
Avoid if
You don't have clear, informed consent from whoever's voice you'd be cloning.
Context
The short 6-second cloning requirement is both the feature and the risk: it makes cloning genuinely easy, which means it's also easy to clone a voice without the speaker's knowledge or consent. Build a consent step into your process before generation, not as an afterthought.
Requirements
This is a self-hosted model; you need a machine capable of running it, with GPU acceleration recommended for practical speed (specific VRAM figures weren't itemized in the pages checked for this review, so test on your actual hardware). It's part of the broader Coqui TTS toolkit (whose code is MPL-2.0 licensed), while the XTTS-v2 model weights themselves carry the separate Coqui Public Model License; check which license actually governs the specific artifact you're using.
Licence & cost
The weights are free to download, but the Coqui Public Model License (CPML) restricts commercial use of generated audio; read the actual license text for your specific use case rather than assuming free-to-download means free-to-use-commercially. You pay for the machine: GPU acceleration is practically necessary for usable speed. Running effort includes building a consent-verification step into any workflow that clones a real person's voice, which is a process cost beyond the compute cost.
First task
Documentation-based suggestion — not run here.
Read first task
Clone a voice with clear consent and test cross-language transfer.
Record (or use) a 6-second sample from someone who has explicitly agreed to have their voice cloned for this test.
Generate speech in the same language as the sample and check how close it sounds.
Generate speech in a different supported language and check whether the vocal character is preserved.
Get the speaker to listen to both outputs and confirm they're comfortable with the result before using it anywhere beyond this test.
Done when you have samples in two languages, both reviewed and approved by the person whose voice was cloned, before any use beyond this test.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on the XTTS-v2 model card, checked 23 September 2026. No voice was cloned and no audio was generated for this review. Exact VRAM requirements were not itemized on the page checked.
XTTS-v2 model card · Sep 23, 2026 — Coqui Public Model License (commercial restrictions), 17 supported languages, 6-second cloning requirement, 24kHz output.
Coqui TTS repository · Sep 23, 2026 — MPL-2.0 licensed toolkit that XTTS-v2 belongs to; active development, streaming latency improvements.
Coqui TTS repository license · Sep 23, 2026 — MPL-2.0 license text governing the toolkit code, distinct from the CPML governing XTTS-v2's model weights.
Further review notes & alternatives
Alternatives and trade-offs
Kokoro wins when you don't need cloning and want a simpler, Apache-2.0 licensed option with a fixed voice list.
OpenAI TTS wins when you need clear commercial usage terms and don't need voice cloning specifically.
F5-TTS wins when expressive delivery matters more than cloning from a short sample; check its own license before use.
Synthesize speech in multiple languages from text using a voice reference.
Run the model locally through the Coqui TTS Python library.
Use optional GPU acceleration when generating speech interactively or at scale.