Boyo Chatterbox Turbo Generate

This is the node that makes Boyonodes' audio stack worth installing. Boyo Chatterbox Turbo Generate takes the model from the Loader, your text, and a reference clip of a voice - as little as five seconds - and synthesizes speech in that voice, complete with emotion tags and a full set of sampling controls. It's the actual TTS engine of the pair, and it's a genuinely impressive model wrapped in a ComfyUI-friendly interface.

How it works

Under the hood it's Chatterbox Turbo running from Hugging Face, with ComfyUI audio plumbing bolted around it. Your reference_audio gets written to a temp WAV and fed to the model as the zero-shot voice-cloning exemplar; the text prompt drives the content; and the sampling parameters - temperature (default 0.8), top_p (0.95), top_k (1000), repetition_penalty (1.2) - steer how the model picks tokens. Optional seed gives you reproducible takes (0 = unseeded), min_p is a stricter sampling floor for cleaner output, and norm_loudness (default on) evens out the final level so your audio doesn't blow the meters.

The one thing that's genuinely fun: emotion tags. The model supports inline tags like [laugh], [chuckle], [sigh], [gasp], [cough], [groan], [sniff], [shush], and [clear throat] - you drop them into the text and the character performs them at that point in the line. This is the difference between a monotone read and a character voice, and it's the feature people build whole pipelines around.

Inputs and outputs

Required: model (from the Loader), text (multiline - the dialogue), reference_audio (the voice sample), and the four sampling params. Optional: seed, min_p, exaggeration, norm_loudness. The single output is audio (AUDIO), ready for a save node, the Duration Analyzer, or the Padder in a lip-sync chain.

Installation

Same as the Loader:

cd ComfyUI/custom_nodes && git clone https://github.com/DragonDiffusionbyBoyo/Boyonodes
pip install chatterbox-tts --no-deps soundfile

restart, and remember the Loader downloads the model on its first run. soundfile is required here (the node needs it to write the reference clip to disk).

Where people get burned

• Voice clone sounds off or robotic - the reference clip quality matters more than length. A clean five-second clip beats a noisy thirty-second one. Speech-to-noise ratio is the lever.
• exaggeration (0–1, default 0) is the emotional-intensity knob, not a distortion effect - crank it if the performance feels flat, not if it sounds bad.
• Repeating the same seed doesn't reproduce a take - the seed only seeds when it's non-zero, and it's applied at generation time; if you changed the reference or text, everything shifts anyway.
• Lip-sync timing - generated speech almost never matches the video length on the first pass. That's what Boyo Audio Padder is for; measure with the Duration Analyzer, pad to target, save.

One more note, same as the Loader article: the loader strips Chatterbox's PerTh authenticity watermark. You're getting voice-clone-quality audio with no provenance marker. Great for game characters and narration; think twice about what else it could be used for. The capability is the model's - the missing watermark is this pack's design choice.