XTTS and Coqui in 2026: Status and Local Alternatives

Short answer: XTTS is still usable in 2026, but it is no longer a safe default. Coqui, the company that produced the XTTS v2 voice-cloning model, is gone, the original repository is unmaintained, and a community fork now carries the code. The model itself did not stop working: XTTS v2 can clone a voice from about six seconds of audio across 17 languages. The real question today is whether a tool from 2023 is still the best pick next to newer local models such as Kokoro, Chatterbox, and Qwen3-TTS.

This article covers what XTTS v2 does, what happened to Coqui, how to run XTTS locally now, and when a newer local model is the better choice. It compares mechanics and licensing rather than sound quality, because quality depends on the voice, the script, and the language you actually feed the model.

What XTTS v2 does, and why people still search for it

XTTS v2 is a text-to-speech model from Coqui that clones a speaker from a short reference clip. The model card lists 17 languages: English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese, Japanese, Hungarian, Korean, and Hindi. It generates 24 kHz audio, supports cross-language cloning (clone in English, then speak Spanish), and adds emotion and style transfer by cloning. The project also advertised streaming with under 200 ms latency.

Those features explain the staying power. XTTS was one of the first open models that made drop-in-a-clip cloning practical without training a voice yourself. Thousands of tutorials, community apps, and pipelines were built around the xtts_v2 checkpoint, and old guides keep sending traffic its way. If you are new to cloning, start with the recording, consent, and quality basics in our voice cloning guide before you compare models.

Coqui the company is gone. The model and code are not.

Coqui, the company behind XTTS, shut down in 2024. The original coqui-ai/TTS repository still exists and is heavily starred, but it is no longer maintained by its original owner. A community fork, idiap/coqui-ai-TTS, now carries the project forward under a new package name, coqui-tts. The fork's README states it is a fork of the original, unmaintained repository, and it notes support for Python 3.10 through 3.14, testing on Ubuntu 24.04, and that PyTorch is no longer bundled from version 0.27.4.

Two licenses travel together here, and they are not the same thing. The code is under MPL-2.0. The XTTS v2 model weights are under the Coqui Public Model License, which the model card links to for the full terms. That license page did not resolve when I checked it, so treat the commercial terms as something you must confirm from the license text itself before you ship paid or client work.

XTTS v2 versus newer local models

XTTS is not the only local option anymore. Kokoro is a small, fast model with a fixed voice set rather than cloning. Chatterbox and Qwen3-TTS added newer cloning pipelines, and both appear in local app catalogs next to other MLX bridge models. The table below is a decision aid, not a quality ranking.

Read the table as a set of constraints. If you need zero-shot cloning in many languages from one clip, XTTS v2 still fits. If you want a small, permissively licensed model for preset narration and you do not need cloning, Kokoro is usually the simpler choice. Use this checklist to decide:

• Choose XTTS v2 if your pipeline already depends on its checkpoints or you specifically need its 17-language clone behavior.
• Choose Kokoro if you want a lightweight, permissively licensed model and fixed voices are enough.
• Choose Chatterbox if you want cloning with either an English-focused or a multilingual variant.
• Choose Qwen3-TTS if you want cloning that accepts a reference transcript and newer controls.
• Check the model license before any paid or client work. The code license and the model license are separate.

Practical steps to run XTTS locally in 2026

You can still run XTTS on a Mac, but plan for a Python setup rather than a double-click app. The maintained fork documents the path. Concrete steps:

1. Install Python 3.10 through 3.14. The fork is tested on that range and will not behave on older interpreters.
2. Install PyTorch, torchaudio, and, for PyTorch 2.9 and later, torchcodec using the official instructions for your platform. These are no longer pulled in automatically.
3. Install the package with uv pip install coqui-tts, or clone the fork and install it editable if you plan to modify code.
4. Prepare a clean reference clip. XTTS v2 was demonstrated with about six seconds of audio, though longer, well-recorded samples usually behave better.
5. Run a short synthesis test before batching a long script. Fix pronunciation and pacing issues in the text, not in the model.
6. Write down the checkpoint, package version, and Python version for every finished project so you can reproduce it later.

The fork also added a caching mechanism for cloned voices, so you can reuse a speaker embedding instead of re-reading the reference clip on every run. That matters when you generate long documents or many short lines from the same voice. If you would rather not maintain a Python environment at all, the local voice studio versus DIY open source comparison walks through that tradeoff.

When a newer local model is the better pick

If a Python environment is not something you want to maintain, a packaged app removes most of this friction. Murmur is a local-first voice studio for Apple Silicon Macs on macOS 15 or later. It bundles Kokoro for fast preset voices and can install optional MLX bridge models for cloning and voice design (these run through MLX), so generation stays on your machine with no per-character fees. It does not ship XTTS, so if your work depends on the XTTS checkpoint specifically, Murmur is not a drop-in replacement.

Be honest about the fit. Murmur is macOS-only. If you need Windows, a browser app, or cloud collaboration, it is the wrong tool. It also does not do live streaming text-to-speech, so real-time voice chat is out. For choosing between local models more broadly, see best local TTS models in 2026, the Kokoro Mac guide, and the Qwen3-TTS status page.

A migration checklist for moving a project off XTTS

XTTS does not need to be abandoned just because it is old. Move only when a specific constraint pushes you. If it does, work through this list:

1. List every finished project that used XTTS and note the reference voice, language, and checkpoint.
2. Pick a replacement per project, not once for everything. A narration job and a character-voice job may need different models.
3. Re-record or re-import reference audio that meets the new model's minimum length and format. Some models need the exact transcript of the reference clip.
4. Regenerate a short sample and compare it against the old output on the same script before committing to a full pass.
5. If you use a permissive model like Kokoro, confirm its license still covers your use case, and keep any attribution it asks for.
6. Archive the old XTTS environment so you can reproduce an earlier deliverable if a client asks for a revision.

Frequently asked questions