⚙️ ChatterBox Official 23-Lang Engine

This is the grown-up ChatterBox. Where the community-model engine loads a separate finetune per language, this one is ResembleAI's official multilingual model - 23 languages baked into a single checkpoint, and you switch between them with a parameter instead of reloading a whole model. If you're doing anything multilingual, this is the ChatterBox you want. It's the same lineage that ResembleAI benchmarked against closed systems like ElevenLabs, it's MIT-licensed, and it landed to a lot of interest when "ChatterBox now supports 23 languages" made the rounds.

Like every engine in the suite, it's a configurator: it outputs a TTS_ENGINE, and the actual speech happens over in 🎤 TTS Text or 📺 TTS SRT.

How it works

One model holds all the languages, so [En:Alice] Hello [De:Hans] Guten Tag [Es:Maria] ¡Hola! in your TTS Text prompt switches language and voice inline without the model-reload stall you get from community finetunes. That's the practical win: language switching is a parameter flip, not a disk load. Zero-shot voice cloning works across all 23 languages from a few seconds of reference audio, and the clone comes from the TTS Text side (via narrator_voice or Character Voices), not from this node.

The inputs and outputs that matter

The two you'll actually set:

• model_version (default v2) - v2 is the current tokenization with the experimental emotion/sound tokens; v1 is the plain original. There are also two community finetunes in the dropdown, Vietnamese (Viterbox) and Egyptian Arabic (oddadmix). Just leave it on v2 for normal use.
• language (default English) - the language to speak. 23 options plus Vietnamese. This is the whole reason you're here.

The generation knobs matter less but exist:

• exaggeration (0–5, default 1) - expressiveness, same idea as classic ChatterBox but a wider range. 1 is a fine baseline.
• temperature (default 0.8) - randomness; keep it near default.
• cfg_weight (default 0.5), repetition_penalty (default 2), min_p (default 0.05), top_p (default 1) - the fine-tuning sampler controls. The 23-Lang model exposes more of these than the community engine; you can mostly ignore them until a specific voice is misbehaving, then nudge repetition_penalty up if it loops or min_p if it rambles.

Single output: TTS_engine, straight into TTS Text or TTS SRT.

Installing it

Part of TTS Audio Suite, so you install the pack. ComfyUI Manager is easiest: search TTS Audio Suite, install, restart - Manager runs install.py, which sorts out the dependency tangle (NumPy, librosa, s3tokenizer) and Python 3.13. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/diodiogod/TTS-Audio-Suite.git, then python install.py inside the folder with your venv active (run the script, don't just pip-install requirements). Linux needs portaudio19-dev libsamplerate0-dev. The official model (~4.3GB) auto-downloads on first use into ComfyUI/models/TTS/chatterbox_official_23lang/, and v1/v2 coexist in that same folder.

Common issues

• The v2 emotion/sound tokens barely do anything. The README is blunt about this: the <laughter>, <whisper>-style special tokens are experimental and often produce "minimal or no audible effects" - ResembleAI hasn't documented them and there's an open upstream issue. Don't build a workflow that depends on them working; treat them as a maybe.
• Perth watermarking. The official model can embed an inaudible watermark for responsible-use tracking. The suite ships it disabled by default for compatibility, but it's worth knowing it exists if you care either way.
• Picking the wrong ChatterBox. If you only ever generate English and want the maximum community-finetune quality for a specific language, the community ⚙️ ChatterBox TTS Engine may suit you better. Use this node when you need multiple languages in one workflow without model-swap latency.
• First run is slow. That's the 4GB+ download, not a hang. Watch the console.