Coqui · Released, not on our provider

XTTS v2 (2026): Coqui’s voice-cloning TTS model, its languages, hardware, paper results and the non-commercial license

XTTS v2 is Coqui’s 2023 open-weight text-to-speech model that clones a voice from a short reference clip in 17 languages at 24 kHz, but its Coqui Public Model License allows only non-commercial use, and this site cannot run it because it is a text chat, not a speech generator.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Qwen3.8 Flash, not XTTS v2.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
Coqui
Type
Text
Released
No date given
OpenRouter id
None

Checked

XTTS v2: facts, with sources

  • The Hugging Face repository coqui/XTTS-v2 was created on October 31, 2023; the model files were uploaded from November 6, 2023, newer checkpoints followed (commits titled "Update to v2.0.3" on November 24 and "Update model with more steps" on December 1, 2023), and the last commit is from December 11, 2023. The model card gives no release date, so we call October 31, 2023 the day the repository was created.Source: huggingface.co · checked 2026-11-19
  • The model card describes XTTS as a voice generation model that clones a voice from a 6-second audio clip, with emotion and style transfer by cloning, cross-language cloning and multilingual generation at a 24 kHz sampling rate. It lists 17 languages: English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese (zh-cn), Japanese, Hungarian, Korean and Hindi. Over XTTS-v1 it adds Hungarian and Korean, better speaker conditioning, several speaker references with interpolation between speakers, and stability and prosody improvements.Source: huggingface.co · checked 2026-11-19
  • Coqui’s own documentation page for XTTS lists 16 languages (the card’s list without Hindi) and says a quick 3-second audio clip is enough to clone a voice, where the card says 6 seconds; the docs also list streaming inference with under 200 ms latency (no hardware is named), fine-tuning support, and cloning from one or several reference files. We could not reconcile the language count and clip length between the card and the docs.Source: tts.readthedocs.io · checked 2026-11-19
  • The XTTS paper by Coqui authors (Interspeech 2024) describes three parts: a VQ-VAE with 13M parameters, a GPT-2-style decoder-only encoder with 443M parameters that predicts the audio codes from text and a conditioning encoder, and a HiFi-GAN decoder with 26M parameters conditioned on a speaker embedding, with output at 24 kHz. The paper says it was trained on 27,281.6 hours across 16 languages, of which English is 14,513.1 hours, for about 2.5 million steps on four NVIDIA A100 80 GB GPUs.Source: arxiv.org · checked 2026-11-19
  • The repository’s file listing shows model.pth at 1,867,929,118 bytes (about 1.87 GB), dvae.pth at 210,514,388 bytes and speakers_xtts.pth at 7,754,818 bytes, plus config.json, vocab.json and mel_stats.pth. The card shows three ways to run it, all through Coqui’s TTS toolkit: the Python API (TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=True), then tts_to_file with a speaker_wav reference and a language), a tts command line with --use_cuda, and loading the checkpoint directly with the Xtts class. Neither the card nor the paper gives a GPU memory, CPU or real-time-factor figure.Source: huggingface.co · checked 2026-11-19
  • Coqui’s own website, as archived on March 3, 2024, showed a single message, "Coqui is shutting down." The model card, however, still describes XTTS as the same or a similar model to what powers Coqui Studio and the Coqui API, and links its demo spaces and community channels.Source: web.archive.org · checked 2026-11-19
  • When we checked on November 19, 2026, coqui.ai/cpml, the license link on the model card, returned a GitHub Pages "Site not found" error, so the license text has to be read in the repository’s own LICENSE.txt (or an archived copy).Source: coqui.ai · checked 2026-11-19 · second-hand
  • The coqui-ai/TTS GitHub repository, the toolkit that runs XTTS, is not archived and its code is under the Mozilla Public License 2.0, which is a different license from the CPML that covers the XTTS-v2 weights; GitHub’s data shows its last push on August 16, 2024.Source: api.github.com · checked 2026-11-19 · second-hand
  • XTTS v2 is not on OpenRouter’s model list (checked November 19, 2026), and this site is a text chat with language models, so it cannot generate speech with it. This site can chat with small text models such as Qwen3.8 Flash, Inkling 2 Small and Hy4 Flash.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

XTTS v2 license: what the Coqui Public Model License allows and what it does not

  • The Coqui Public Model License 1.0.0 in the repository’s LICENSE.txt says it allows "only non-commercial use of a machine learning model and its outputs". Non-commercial purposes are personal use for research, experiment, testing, study, private entertainment, hobby projects, amateur pursuits or religious observance; use by for-profit entities for testing, evaluation or non-commercial research and development; and use by charities for charitable purposes or testing, in each case only so far as you receive no direct or indirect payment arising from the use of the model or its output.Source: huggingface.co · checked 2026-11-19
  • The same license says use for revenue-generating activity, including projects directly funded by government grants, is not a non-commercial purpose, and neither is using the model to train other models for commercial use. Anyone who gets the model, a modification or their output from you must also get the terms or their URL; the licenses cannot be sublicensed or transferred; they end at once if you claim in writing that the model infringes a patent; after a first written notice of a violation you have 30 days to come into compliance before they end; and the model and its output come without warranty.Source: huggingface.co · checked 2026-11-19

License

XTTS v2 license (Coqui Public Model License 1.0.0) in plain words

The license is the first thing to check with XTTS v2: it covers the weights and their output, and it allows only non-commercial use. A company may test and evaluate it, but anything that earns money, including a grant-funded project, falls outside it. If your plan needs commercial rights, look at a model with an Apache 2.0 license such as ZONOS3-Mini or the smaller Kitten TTS 2 Mini, and read their licenses too.

Our table: a plain-words summary of the license text in the XTTS-v2 repository (LICENSE.txt), read on November 19, 2026; this is not legal advice, read the license itself before you rely on it · checked 2026-11-19
UseVerdictDetail
Personal research, testing and studyAllowedExperiment, testing, personal study; only while you receive no direct or indirect payment from the model or its output
Hobby, private entertainment, religious observanceAllowedHobby projects, amateur pursuits and private entertainment, under the same no-payment condition
A for-profit companyTesting onlyTesting, evaluation or non-commercial research and development; training other models for commercial use is not allowed
A charityAllowed for charitable purposesCharitable purposes, testing or evaluation; revenue-generating activity is not covered
Revenue-generating activityNot allowedIncludes projects directly funded by government grants
Passing it onWith the termsAnyone who gets the model, a modification or their output from you must get the terms or their URL; you cannot sublicense or transfer your licenses
If you break the terms30 days to fixAfter the first written notice you have 30 days to come into compliance and correct past violations, otherwise the licenses end
Patent claims against the modelLicenses endA written claim that the model infringes a patent ends your licenses at once
WarrantyNoneThe model and its output come as is, without warranty, and the licensor is not liable for damages

Source: huggingface.co

Our table

XTTS v2 English results in the paper: error rate, naturalness and speaker similarity against other models

  • In the paper’s English test (240 FLORES+ sentences, references from 20 DAPS speakers, character error rate by Whisper large-v3, naturalness by UTMOS, speaker similarity by ECAPA2 cosine similarity), XTTS scores a CER of 0.5425, a UTMOS of 4.007 and a speaker similarity of 0.6423, beside Tortoise (1.0934, 4.0883, 0.5492) and HierSpeech++ (0.7741, 4.457, 0.6530). In listener preference tests the authors report XTTS ahead of HierSpeech++ and Mega-TTS 2 on naturalness, acoustic quality and human likeness (comparative score 0.41 and 0.92 on a scale of -2 to +2) and slightly behind on speaker similarity (-0.31 and -0.39). These are the authors’ own results and we did not run them; the paper links the repository’s v2.0.2 tag, so they may not describe the current weights exactly.Source: arxiv.org · checked 2026-11-19

Paper figures

XTTS v2 English test from the XTTS paper: XTTS against other zero-shot voice models

The paper is from mid-2024 and compares XTTS with the open models of the time, so it says nothing about 2026 models such as ZONOS3-Mini. Its English row shows a very low error rate for XTTS, with naturalness and speaker similarity in the middle of the field; the authors themselves note that HierSpeech++ scored better on naturalness. Read the table as a snapshot of 2023-2024 open voice cloning.

Our table: Table 2 of the XTTS paper (arXiv 2406.04904), read on November 19, 2026; character error rate (lower is better), UTMOS naturalness and speaker similarity (higher is better), authors’ own results that we did not run · checked 2026-11-19
ModelCharacter error rateUTMOSSpeaker similarity
Ground truth recordingsnot given4.2775 ± 0.150.8952
XTTS (the paper’s Exp. 3)0.54254.007 ± 0.250.6423
Tortoise1.09344.0883 ± 0.310.5492
StyleTTS 20.67894.4260 ± 0.070.4728
Mega-TTS 21.42694.184 ± 0.170.6428
HierSpeech++0.77414.457 ± 0.060.6530
Original YourTTS2.87363.6034 ± 0.290.4621

Source: arxiv.org

Our table

XTTS v2 languages: error rate and speaker similarity in the 16 languages of the paper

  • Across the 16 languages of the paper’s multilingual test, XTTS averages a character error rate of 2.0646 and a speaker similarity of 0.5046, against 4.7949 and 0.4704 for YourTTS trained on the same data; English is lowest at 0.5425 and Japanese (5.3748), Chinese (5.2016) and Korean (4.0647) are highest. The paper’s own table has YourTTS ahead on error rate in German and Spanish, and the authors say XTTS has better error rate and speaker similarity in almost all languages. The paper evaluates 16 languages, not the 17 on the card.Source: arxiv.org · checked 2026-11-19

By language

XTTS v2 per-language character error rate and speaker similarity, with YourTTS trained on the same data

The card’s language list is longer than the paper’s test, so Hindi has no printed score at all. The scores spread widely: the paper’s English error rate is about a tenth of its Japanese and Chinese ones. For languages beyond these sixteen, a model with a wider list such as ZONOS3-Mini, which names 43, may fit better. If all you need is text, you can chat with a small model here.

Our table: Table 4 of the XTTS paper (arXiv 2406.04904), read on November 19, 2026; character error rate (lower is better) and speaker similarity (higher is better); the paper has no row for Hindi, which the model card lists · checked 2026-11-19
LanguageXTTS character error rateXTTS speaker similarityYourTTS error rate
Arabic (ar)3.35030.500711.1713
Czech (cs)1.32950.46554.0174
German (de)3.16940.51752.2411
English (en)0.54250.64232.9727
Spanish (es)1.46060.53711.0926
French (fr)1.49370.47993.3965
Hungarian (hu)1.46220.45704.5098
Italian (it)0.79820.50081.7010
Japanese (ja)5.37480.520710.2808
Korean (ko)4.06470.47608.8567
Dutch (nl)0.9460.48253.4228
Polish (pl)0.75930.48331.5925
Portuguese (pt)1.10680.50331.5481
Russian (ru)0.9320.50122.8566
Turkish (tr)1.0420.50312.6367
Chinese (zh-cn)5.20160.502314.4220
Average of the 162.06460.50464.7949

Source: arxiv.org

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to XTTS v2 you can use now

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

Inkling 2 Small

Thinking Machines Lab

$0.30 / $1 (per million tokens)

  • Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
  • A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
  • Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14

Hy4 Flash

Tencent

$0.04 / $0.14 (per million tokens)

  • Tencent’s Hy4 flash announcement, dated November 16, 2026, presents it as the newest member of the Hy model family. Tencent recommends it for most everyday tasks (deep research, document processing, PowerPoint generation, simple front-end work, computer use) and points complex tasks such as long-horizon software engineering to Hy4. Tencent Cloud’s API guide calls it "a leaner, more efficient everyday workhorse".Source: tencent.com · checked 2026-11-17
  • The Tencent announcement is dated November 16, 2026, and Tencent Cloud’s model, pricing and API pages for hy4-flash were last updated the same day. The announcement says only that Hy4 flash is available now; it gives no separate release date or version name.Source: tencent.com · checked 2026-11-17
  • Tencent says Hy4 flash has 224B total parameters (including 64B Engram parameters), 12B active parameters, a context of up to 1M tokens and multimodal input; the announcement does not state a maximum output or the output types. It says the KV cache is 1/6 the size of Hy4’s and that decoding exceeds 150 tokens per second. These are the maker’s own figures, not tested by us.Source: tencent.com · checked 2026-11-17

FAQ

XTTS v2 FAQ

What is XTTS v2?

XTTS v2 is a multilingual voice-cloning text-to-speech model from Coqui. Its Hugging Face card says it clones a voice from a 6-second clip, transfers emotion and style by cloning, works across languages and outputs 24 kHz audio in 17 languages. The repository was created on October 31, 2023, and its last commit is from December 11, 2023.

Can I use XTTS v2 commercially?

Not under its license. The Coqui Public Model License 1.0.0 allows only non-commercial use of the model and its outputs: personal research, testing, hobby use, for-profit testing and evaluation, and charity use. Revenue-generating activity, including grant-funded projects, is not covered, and neither is training other models for commercial use. This is our summary, not legal advice; read LICENSE.txt in the repository.

What happened to Coqui, and is XTTS v2 still supported?

Coqui’s own website, as archived on March 3, 2024, said the company was shutting down, and the card’s license link, coqui.ai/cpml, returned a "Site not found" error when we checked on November 19, 2026. The coqui-ai/TTS toolkit repository is not archived, but GitHub shows its last push on August 16, 2024, and the model repository’s last commit is from December 11, 2023.

What hardware does XTTS v2 need, and how fast is it?

The main weight file, model.pth, is about 1.87 GB, and the card runs the model through Coqui’s TTS toolkit with a GPU switch (gpu=True or --use_cuda). Coqui’s docs claim streaming with under 200 ms latency but name no hardware, and neither the card nor the paper gives a memory or real-time-factor figure, so we cannot say what GPU you need. The paper trained it on four NVIDIA A100 80 GB GPUs.

How good is XTTS v2?

By the paper’s own English test, which we did not run, XTTS reaches a character error rate of 0.5425, a UTMOS of 4.007 and a speaker similarity of 0.6423, and across 16 languages its average error rate is 2.0646. It beats Tortoise on error rate and trails HierSpeech++ on naturalness. These figures date from 2024, come from the v2.0.2 checkpoint the paper links rather than necessarily the current weights, and do not include newer models such as ZONOS3-Mini.

Does XTTS v2 clone voices, and what about consent?

Yes, that is its main feature: it clones from a short clip of one or several recordings. Use only recordings of people who have agreed, do not present generated speech as a real person, and check the rules where you live. The license does not mention consent and this note is ours, not Coqui’s.

Can I generate speech with XTTS v2 here?

No. This site is a text chat with language models, so it cannot produce audio with XTTS v2 or any other speech model, and the model is not on OpenRouter’s list. For text, a typical reply to Inkling 2 Small costs about 1 credit on this site (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Hy4 Flash is another small model.

Is this the official Coqui website?

No. This page is unofficial and not affiliated with or endorsed by Coqui. Facts link to the pages we read.