AI voice model · Coqui

Coqui TTS and XTTS v2: The Open-Source Voice Toolkit, Explained

Coqui TTS is a free, MPL-2.0 Python toolkit for text to speech, and its best-known model, XTTS v2, clones a voice from a few seconds of audio in 17 languages for non-commercial use. If you need commercial voiceovers without running Python, SupaVocal is a hosted alternative from $9 for the first month.

  • Toolkit: open source, MPL-2.0
  • XTTS v2: 17 languages, 24 kHz
  • SupaVocal: no setup, commercial use
Start creating voiceovers
17
Languages in XTTS v2
6 s
Reference clip the XTTS v2 model card asks for
24 kHz
XTTS v2 output sample rate
0.28.0
Latest coqui-tts release, October 2026

What is Coqui TTS?

Coqui TTS is an open-source deep learning library for generating speech from text, written in Python on PyTorch. It started at Coqui, a Berlin start-up, and bundles ready-to-use pretrained models, training recipes for new voices and languages, and dataset tools in one package. The code is released under the Mozilla Public License 2.0.

Inside one install you get many architectures: spectrogram models such as Tacotron 2, Glow-TTS and FastPitch, end-to-end models such as VITS, YourTTS, Tortoise, Bark and XTTS, a family of vocoders such as HiFiGAN, and voice conversion models such as FreeVC. Through the Fairseq MMS models it covers pretrained voices in about 1,100 languages.

The company itself is gone. In early 2024 the Coqui website read "Coqui is shutting down", and the original coqui-ai/TTS repository stopped at release 0.22.0 in December 2023. Development continues in a community fork, published on PyPI as `coqui-tts`, which is the version to install today.

Coqui TTS specs at a glance

The maintained coqui-tts package and its XTTS v2 model as of November 2026.

Original makerCoqui (Berlin), shut down in early 2024
Maintained byThe idiap fork of the original repository, PyPI package coqui-tts
Latest releasecoqui-tts 0.28.0, released October 19, 2026
Python3.10 to 3.14, with PyTorch 2.2 or newer installed separately
PlatformsPrebuilt wheels for Linux, macOS and Windows; tested on Ubuntu 24.04
Code licenseMPL-2.0
Flagship modelXTTS v2, multilingual voice cloning built on Tortoise
XTTS v2 languages17: English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese, Japanese, Hungarian, Korean, Hindi
XTTS v2 audio24 kHz output, streaming with under 200 ms latency
XTTS v2 licenseCoqui Public Model License: non-commercial use of the model and its outputs only
Other modelsVITS, YourTTS, Tortoise, Bark, Tacotron 2, Glow-TTS, Fairseq MMS voices in about 1,100 languages
Voice conversionFreeVC, OpenVoice and kNN-VC

Original coqui-ai/TTS vs the maintained coqui-tts fork

Two Python packages share the Coqui TTS name. Install the fork: the original no longer gets fixes for new Python or PyTorch releases.

coqui-tts (fork)TTS (original)
Installpip install coqui-ttspip install TTS
StatusMaintainedUnmaintained since Coqui closed
Latest release0.28.0, October 20260.22.0, December 2023
Python3.10 to 3.143.9 to 3.11
PyTorchInstalled by you (from 0.27.4)Pulled in automatically
Windows and macOS wheelsYes, since 0.24.2No prebuilt wheels
Cached cloned voicesYes, since 0.27.0No
Extra voice conversionOpenVoice, kNN-VCFreeVC only
Import namefrom TTS.api import TTSfrom TTS.api import TTS

Both packages import as TTS, so install only one of them in a given Python environment.

What Coqui XTTS and the toolkit do well

These strengths explain why Coqui TTS is still one of the most used open voice toolkits years after the company closed.

Cloning from seconds

XTTS v2 copies a voice from a short reference clip with no training run, and accepts several clips of one speaker for a closer match.

Cross-language cloning

Clone a voice from an English clip and make it speak any of the other 16 XTTS v2 languages.

Streaming output

XTTS can stream its first audio chunk in under 200 ms, which makes it usable for voice assistants.

Preset speakers

XTTS v2 ships named studio speakers, such as Ana Florence, for when you have no reference clip.

Train your own

Recipes, a Trainer API and dataset tools let you fine-tune XTTS or train a VITS voice in a new language.

One API, many models

The same TTS() call loads XTTS, VITS, Bark, Tortoise or a Fairseq voice, so you can compare models in a few lines.

How to run Coqui TTS locally

The quickest route to Coqui TTS is the maintained coqui-tts package. These are the steps its README gives.

  1. 1

    Install PyTorch first

    From release 0.27.4, PyTorch is no longer bundled. Install torch and torchaudio for your CPU, CUDA or ROCm setup (plus torchcodec with PyTorch 2.9 or newer) on Python 3.10 to 3.14.

  2. 2

    Install the package

    Run pip install coqui-tts, or uv pip install coqui-tts as the maintainers recommend. Add extras such as coqui-tts[server,ja] for the demo server or Japanese text.

  3. 3

    Generate speech

    In Python, load TTS("tts_models/multilingual/multi-dataset/xtts_v2"), then call tts_to_file() with your text, a language code and either a preset speaker or a speaker_wav clip. The weights download on first use.

  4. 4

    Or use the command line

    Run tts --list_models to see every model, then tts --text "..." --model_name <name> --out_path speech.wav. For XTTS, add --speaker_wav and --language_idx.

Coqui TTS install and deployment options

Pick the route by who will call the model and where it has to run.

RouteRuns onBest forNotes
coqui-tts Python APICPU or CUDA GPUScripts, notebooks, batch narrationLowest-level control; Synthesizer API adds timestamps
tts command lineCPU or CUDA GPUQuick tests, shell pipelines--use_cuda for GPU; --pipe_out streams the WAV
Docker imageCPU image, GPU images documentedTrying it without a Python installghcr.io/idiap/coqui-tts-cpu, demo server on port 5002
Demo serverAny machine running the packageA local web page or HTTP endpointInstall the server extra; pick a model with --model_name

XTTS v2 splits long text into sentences by default. Turning that off gives smoother joins but uses more GPU memory and can hit the model's context limit.

Coqui voice cloning with XTTS v2

Coqui voice cloning runs without any training: you pass audio, and XTTS v2 conditions its output on that speaker.

  • Single reference. Pass one clean clip as speaker_wav. The XTTS v2 model card asks for about 6 seconds; the newer fork documentation says 3 seconds can work.
  • Multiple references. Pass a list of clips, or every WAV in a folder, for a closer match. Runtime does not grow with the number of clips.
  • Cached voices. From coqui-tts 0.27.0 you can store a cloned voice under a custom speaker ID and skip re-sending the audio.
  • Voice conversion. FreeVC, OpenVoice and kNN-VC turn one recording into another speaker's voice, and tts_with_vc_to_file() adds a cloned voice to any single-speaker model.

The catch is the license. XTTS v2 weights are under the Coqui Public Model License, which allows only non-commercial use of the model and its outputs. A monetized YouTube voiceover or a paid audiobook made with XTTS v2 falls outside it. Clone only your own voice or one whose owner has agreed.

Coqui TTS limitations to plan for

None of these stop a hobby project, but each one matters before you build a product on Coqui TTS.

Do

  • Check the license of each model you ship: the toolkit is MPL-2.0, XTTS v2 weights are non-commercial
  • Install the maintained coqui-tts package, not the old TTS release
  • Install PyTorch yourself and match it to your CUDA or CPU setup
  • Give XTTS v2 a clean, single-speaker reference clip

Don't

  • Expect fixes or new models in the original coqui-ai/TTS repository
  • Use XTTS v2 output in paid or monetized work
  • Ask XTTS v2 for a language outside its 17 without fine-tuning
  • Clone a voice without the speaker's consent

Which Coqui TTS model should you use?

The models most people load from Coqui TTS, by what each is for.

ModelTypeVoice cloningLanguagesUse it for
XTTS v2End-to-end, built on TortoiseYes, from a short clip17Multilingual cloning and streaming
YourTTSEnd-to-endYes, from a clipEnglish, French, PortugueseLighter zero-shot cloning
VITSEnd-to-endThrough voice conversionPer checkpoint, such as English VCTKFast single or multi-speaker voices
Fairseq MMSVITS checkpointsThrough voice conversionAbout 1,100Rare languages
TortoiseEnd-to-endYesEnglishSlow, high-quality English
BarkEnd-to-endYesMultilingualExpressive and non-speech sounds

Each checkpoint carries its own license. The MPL-2.0 license covers the Coqui TTS code, not the weights you download through it.

A Coqui TTS alternative for commercial voiceovers

Coqui TTS is the right choice if you want to self-host, train a voice in a new language, or need XTTS v2's 17 languages for personal and research work. It costs nothing but your hardware time.

If you need voiceovers you can sell, and don't want to manage Python, PyTorch and model licenses, SupaVocal is a separate, hosted AI voice studio; it does not run Coqui TTS, XTTS or any Coqui model. You paste up to 15,000 characters into text to speech, pick one of 30+ English and Chinese voices and download the audio, with commercial use on every plan.

Plans start at $19 a month, $9 for the first month, for 200,000 characters, about 200 minutes of speech. See pricing. SupaVocal speaks fewer languages than XTTS v2, so test yours first.

Frequently asked questions

What is Coqui TTS?

Coqui TTS is an open-source Python library for text to speech, first built by the Berlin start-up Coqui. It bundles pretrained models such as XTTS v2, VITS, YourTTS, Tortoise and Bark, training recipes and dataset tools. The code is MPL-2.0. Since Coqui closed in early 2024, a community fork maintains it as the coqui-tts package.

Is Coqui TTS still maintained?

The original coqui-ai/TTS repository is not: its last release, 0.22.0, came out in December 2023. A fork published on PyPI as coqui-tts carries on, with release 0.28.0 in October 2026, support for Python 3.10 to 3.14, prebuilt Windows and macOS wheels and new voice conversion models. Install that one.

Is Coqui XTTS v2 free for commercial use?

No. The Coqui TTS code is MPL-2.0, but the XTTS v2 weights are under the Coqui Public Model License, which allows only non-commercial use of the model and of the audio it produces. Paid products, monetized videos and client work fall outside it. Other models in the toolkit carry their own licenses, so check each one.

How many languages does XTTS v2 support?

XTTS v2 supports 17 languages: English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese, Japanese, Hungarian, Korean and Hindi. Hungarian and Korean were added over XTTS v1. It can also clone a voice from a clip in one language and speak it in another.

How long a clip does Coqui voice cloning need?

The XTTS v2 model card asks for a clip of about 6 seconds, and the maintained fork's documentation says 3 seconds can be enough. Clean, single-speaker audio matters more than length. You can pass several clips of the same speaker for a closer match without slowing generation down.

How do I install Coqui TTS?

Install PyTorch and torchaudio for your hardware first, then run pip install coqui-tts on Python 3.10 to 3.14. Load XTTS v2 with TTS("tts_models/multilingual/multi-dataset/xtts_v2") and call tts_to_file with your text, a language code and a reference clip. A CPU Docker image is available if you want to skip the install.

Does SupaVocal use Coqui TTS?

No. SupaVocal is a separate, hosted voice studio and does not run Coqui TTS, XTTS v2 or any Coqui model. It is an alternative for people who want text to speech, voice cloning and voice design without a Python setup: 30+ English and Chinese voices, commercial use on every plan, from $19 a month with the first month at $9.

Is there an online alternative to Coqui TTS?

Yes. If self-hosting Coqui TTS is more setup than you need, or XTTS v2's non-commercial license rules it out, SupaVocal runs in the browser: paste a script of up to 15,000 characters, choose a voice or clone one from 3 seconds of audio, and download the result. Failed generations refund their credits automatically.

Skip the Python setup: try SupaVocal instead of Coqui TTS

Type a script, pick from 30+ English and Chinese voices or clone your own from a 3-second clip. $9 for your first month, commercial use included.

Start creating voiceovers