ChatterBox TTS Loader (Manual)

ChatterBox TTS Loader (Manual) is the bring-your-own-files loader for the pack's voice cloning nodes. It doesn't download anything - you put the ChatterBox model weights in ComfyUI/models/chatterbox/ yourself and point this node at each piece. It's the most control you get over a ChatterBox setup: you pick the exact checkpoint files, the tokenizer, the device, and (optionally) the built-in voice conditionals. If you'd rather not manage files at all, the Auto loader exists precisely for that; this one is for people who already have the weights or want a specific version.

How it works

ChatterBox's TTS is a three-piece stack, and this loader assembles it. The s3gen.safetensors file provides the speech-generation (S3Gen) model; t3_cfg.safetensors is the text-to-token (T3) model; ve.safetensors is the voice encoder; tokenizer.json is the text tokenizer. The loader instantiates each model, loads the weights, moves everything to your chosen device, and packages it into a CHATTERBOX_TTS object with a fixed sample rate and the pieces wired together. The optional conds.pt file loads pre-built voice conditionals for ChatterBox's built-in voices. You'll notice the code loads T3 with strict=False - different checkpoints have different vocab sizes, so it tolerates mismatches and warns rather than dying.

The inputs

• s3gen - pick s3gen.safetensors from models/chatterbox/
• t3_cfg - pick t3_cfg.safetensors
• ve - pick ve.safetensors
• tokenizer - pick tokenizer.json
• device - auto (best available), cuda, or cpu
• conds (optional) - conds.pt for built-in voices

Output is a single model (CHATTERBOX_TTS) that feeds ChatterBox TTS (Manual) - not the Simple node, which expects the Auto loader's model object.

Install and getting files

Pack install via ComfyUI Manager (search "Latent Astronaut Suite") or:

cd ComfyUI/custom_nodes
git clone https://github.com/latentastronaut/comfyui-latent-astronaut-suite

Restart. install.py installs chatterbox-tts with --no-deps plus librosa on first boot. Then the manual part: the drop-downs only list files already inside ComfyUI/models/chatterbox/ (create it if it doesn't exist), so that's where the weights go. s3gen.safetensors, t3_cfg.safetensors, and ve.safetensors are large downloads from the model's official release - if that's friction, use the Auto loader and let it fetch from HuggingFace instead.

Where people get burned

Missing files are the whole failure mode - load the node and any dropdown showing just "none" means that piece isn't in the folder. Keep all four required files together and refresh after copying them in. And as with the whole ChatterBox family, the watermarker is intentionally disabled in this pack (it segfaults), so don't hunt for a missing watermark as a sign of breakage.