Chatterbox Multilingual v3: Free Voice Cloning in 25 Languages, Self-Hosted

Chatterbox Multilingual v3 is Resemble AI's open-source text-to-speech model, released June 10, 2026. It clones voices from a short reference clip and speaks 25 languages (21 languages plus 4 dialects) on the same 0.5B Llama backbone as v2. It's MIT-licensed, runs self-hosted on modest hardware, and embeds a PerTh watermark in every output by default.

That last sentence is the part most write-ups skip. This guide covers the release, the Turbo-vs-Multilingual decision, a verified self-host route, and the watermark — honestly.

---

What Is Chatterbox Multilingual v3? {#what-is-chatterbox-v3}

Chatterbox is Resemble AI's open-source TTS family. The Multilingual v3 release on June 10, 2026 is the current flagship for anything beyond English:

The pitch is simple: commercial voice cloning quality, in 25 languages, that you can run on your own machine under a license that lets you ship products with it. Until recently, that combination meant paying ElevenLabs per character or wrestling with the license terms around XTTS v2. Chatterbox's MIT license removes the legal ambiguity.

The model is small by 2026 standards. At 0.5B parameters (and 350M for Turbo), Chatterbox runs on modest GPUs and even CPU — this is not a "you need a 4090" situation. We're deliberately not quoting specific VRAM or latency numbers here because the ones floating around are unverified; the community server below supports CUDA, ROCm, and CPU-only installs, so start with whatever you have.

---

What "25 Languages" Actually Means {#languages}

Resemble counts 25 total languages: 21 languages plus 4 dialects. On top of that, they publish 6 single-language "Language Pack" variants — dedicated single-language checkpoints for teams that only need one language.

Two practical notes:

1. The dialects count toward the 25. If you see "21 languages" and "25 languages" quoted in different places, both are right — the difference is the 4 dialects.
2. Language Packs are separate model files, not a toggle. If you're deploying one language in production, a Language Pack is the purpose-built option.

For how Chatterbox's output quality stacks up against the other local engines in practice, see our Kokoro vs XTTS vs Chatterbox comparison.

---

Chatterbox Turbo vs Multilingual v3 {#turbo-vs-multilingual}

Resemble now ships two distinct open models, and picking the wrong one is the most common setup mistake.

Chatterbox-Turbo (December 2025) is a 350M-parameter, MIT-licensed, English-only model. Its distinguishing feature is native paralinguistic tags — you can write [laugh], [cough], or [chuckle] directly into the text and the model performs them. It also does zero-shot voice cloning from a few seconds of audio. Resemble claims sub-150ms first audio and quality that beats ElevenLabs and Cartesia — both are vendor claims we haven't independently verified. What is confirmed (via the Hugging Face discussions): Turbo is English-only, and multilingual users are pointed to Multilingual v3.

The decision is one question: do you need anything other than English? If no, and you want speed plus expressive tags, run Turbo. If yes — even one non-English language — Turbo can't help you; run Multilingual v3 or a Language Pack.

---

The PerTh Watermark Is On by Default {#perth-watermark}

Here's the part you should know before building anything: every self-hosted Chatterbox output carries a PerTh audio watermark, embedded by default. This isn't a cloud-only policy — it's baked into the open-source releases.

Resemble states the watermark survives MP3 and Opus compression, telephony codecs, clipping, and resampling. In other words, it's designed to persist through the normal things that happen to audio on its way to a listener. It's inaudible; you won't hear it, and your users won't either.

Why does it exist? Provenance. Chatterbox clones voices from seconds of audio under an MIT license — a combination that's easy to abuse. The watermark means synthetic audio can be identified as synthetic after the fact. That's arguably why Resemble can release this openly at all.

What it means for you:

• Podcasts, videos, apps, IVR: no practical impact. The watermark is inaudible and doesn't degrade quality in any way you'll notice.
• Products where clients require watermark-free masters: know it's there and disclose it. Don't discover this after delivery.
• Detection: the flip side is a feature — if someone clones a voice with Chatterbox to impersonate a person, the output is traceable as synthetic.

We're covering this because most Chatterbox tutorials don't. It's on by default, and defaults matter.

---

Self-Hosted Setup {#setup}

The route most people use is the community devnen/Chatterbox-TTS-Server project: a FastAPI-style server with a web UI, an OpenAI-compatible API, and install paths for NVIDIA CUDA, AMD ROCm, and CPU-only. These commands come from the project's README as of July 2026.

1. Clone and create a venv (Linux/macOS shown):

git clone https://github.com/devnen/Chatterbox-TTS-Server.git
cd Chatterbox-TTS-Server
python3 -m venv venv
source venv/bin/activate

On Windows PowerShell, the venv lines are python -m venv venv then .\venv\Scripts\activate.

2. Install for your hardware — pick ONE requirements file:

pip install --upgrade pip

# CPU only:
pip install -r requirements.txt

# NVIDIA GPU (CUDA 12.1 — RTX 20/30/40 series):
pip install -r requirements-nvidia.txt

# NVIDIA GPU (CUDA 12.8 — RTX 5090 / Blackwell):
pip install -r requirements-nvidia-cu128.txt

# AMD ROCm (Linux):
pip install -r requirements-rocm-init.txt
pip install -r requirements-rocm.txt

3. Install the Chatterbox package, then start the server:

pip install --no-deps git+https://github.com/devnen/chatterbox-v2.git@master
python server.py

(The repo also ships launchers: ./start.sh on Linux/macOS, start.bat on Windows.)

4. Open the web UI at http://localhost:8004 (API docs at http://localhost:8004/docs).

The server exposes an OpenAI-compatible API at /v1/audio/speech and /v1/audio/voices, so anything that speaks the OpenAI TTS API can point at your box instead — pairing Chatterbox with a local STT model gets you a full local speech-to-speech assistant. The UI includes an engine selector for the Multilingual model with a language dropdown. Prefer containers? The repo provides Docker Compose files for NVIDIA (docker compose up -d --build), ROCm (docker compose -f docker-compose-rocm.yml up -d --build), and CPU (docker compose -f docker-compose-cpu.yml up -d --build).

If you're comparing setup effort across engines first, our Kokoro TTS local setup guide covers the lightest-weight alternative, and the best local TTS models roundup maps the whole field.

---

Voice Cloning Ethics: Get Consent {#ethics}

Chatterbox clones a voice from a few seconds of audio. That's the feature — and the risk.

The rule is short: cloning a real person's voice requires their consent. Not "they posted a podcast so it's public." Consent. Cloning your own voice, a voice actor you've contracted, or a synthetic voice you designed is fine. Cloning a colleague, a celebrity, or a family member without asking is not — and depending on your jurisdiction it may be illegal, MIT license or not.

The default-on PerTh watermark exists partly for exactly this: provenance. If cloned audio surfaces somewhere it shouldn't, it can be identified as synthetic. Build with that assumption rather than against it.

---

Verdict {#verdict}

Chatterbox Multilingual v3 is the current default recommendation for self-hosted voice cloning beyond English. 25 languages, MIT license, a 0.5B model that runs on modest hardware, and a mature community server with an OpenAI-compatible API — that combination didn't exist a year ago at this quality level.

Choose it if: you need multilingual voice cloning, you want to self-host, and you want a license you can build a product on.

Choose Chatterbox-Turbo instead if: you're English-only and want the expressive tags and the (vendor-claimed) low latency.

Look elsewhere if: you need guaranteed watermark-free output — PerTh is on by default and that's the deal — or you only need clean English narration without cloning, where smaller models like Kokoro are cheaper to run. See how they compare in our three-way TTS comparison.

---

Sources {#sources}

• Resemble AI: Chatterbox Multilingual v3 announcement — release details, language count, watermarking
• Resemble AI on Hugging Face — model cards and community discussions
• devnen/Chatterbox-TTS-Server README — install commands verified July 2026

---

FAQ {#faq}