Chatterbox · Text to Speech

Chatterbox is the open TTS model that made local text-to-speech competitive with paid services - a 0.5B model on a Llama backbone, MIT-licensed, zero-shot voice cloning from a few seconds of reference audio, and an emotion dial that was genuinely unusual in an open model when it landed in May 2025.

Chatterbox · Text to Speech puts that model behind an API node for 5.4 credits a run. Which raises the obvious question: why pay, when the weights are MIT and run under 8 GB? If you've already got the stack working, you shouldn't. If you don't want a TTS node pack, a tokenizer quirk and a reference-audio pipeline in your life for one narration job, five cents is a bargain.

What it is

resemble-ai/chatterbox-v1/text-to-speech on PopNow's servers, wrapped as a ComfyUI node. Text in, AUDIO out - a normal ComfyUI audio value, so Save Audio writes it and anything downstream that takes audio takes it. No GPU, no weights, no CUDA context. The audio-generation essay rates Chatterbox as the open quality reference (100 threads in the corpus, and the one people compare to ElevenLabs), with F5-TTS as the faster, worse-sounding alternative.

The inputs

text is what gets spoken, and the tooltip is unusually strict: up to 1,000 characters of English, read exactly as written, tags read aloud. So there's no SSML, no pause markers, no pronunciation control. Spell out anything the model might mangle - an acronym, a number you want read as a year, a name.

This is the node's biggest limit and worth planning around. Long narration has to be split and stitched, and chunking at sentence boundaries keeps the voice consistent enough that listeners don't hear the seams.

exaggeration (0.5) is the leftover from the open model's original selling point: 0 flat, 1 dramatic. It's the reason people pick Chatterbox family models - at 0.3 you get a neutral read, at 0.7 you get something with eyebrow movement. For narration, 0.2–0.4. For character dialogue, push it up and listen.

temperature (0.7) is variation between takes - the conventional sampling temperature, 0.05 to 2. Low is repeatable and a little flat; high is more expressive and occasionally unhinged.

guidance_scale (0.5) is not CFG. Here it's "how closely the reading follows the text's natural pacing", 0.1 to 1. If you've arrived from image models and you're looking for a CFG dial, this isn't it - and cranking it won't make the voice follow your prose harder.

reference_audio_url (optional, an AUDIO socket) is voice cloning: feed a short clip and the reading takes on its timbre. Note what it doesn't do - it doesn't transfer acting, and it doesn't fix a bad read. A monotone reference gives you a monotone voice.

Outputs: audio and task id. Several results come back as a batch, and Save Audio writes one file each.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Or Manager → search PopNow → Install → restart. Key from popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY=sk-popnow-... python main.py.

No dependencies - the pack uses the aiohttp/torch/Pillow/PyAV ComfyUI already ships, so there's no Chatterbox install, no model download, no CUDA extension build. That's the entire point of the node, honestly.

Where people get burned

English only. Other languages are on the Multilingual line of the model family, and this node is the base English model. The pack does carry speech nodes for other languages (Qwen3 TTS, MiniMax Speech, Seed Speech) - pick one of those instead of feeding French into this and wondering why it sounds drunk.

1,000 characters, hard. Splitting is on you. And because each chunk is a separate billed run, a five-minute narration is a dozen runs and sixty-odd credits - still six cents, but plan the split before you queue.

Cloning is a rights question, not a technical one. Zero-shot cloning is exactly as easy as it sounds, and a hosted service still applies its own moderation - this is the one node in the pack where "it's just an API call" doesn't insulate you from consequences.

No key, no noise. The node fails loudly with No PopNow API key is set rather than silently producing silence, which is the failure mode you want.