VoiceBridge TTS Loader

The "voice" half of VoiceBridge runs on Qwen3-TTS, Alibaba's open text-to-speech model, and this is the node that loads it. Qwen3-TTS is worth knowing about outside this pack, too: it's the model behind a decent slice of the TTS chatter in the image/video generation community (a few hundred threads by mid-2026), and its big selling points are zero-shot voice cloning from a few seconds of reference audio and support for the major languages. This loader is how you get that model into a ComfyUI graph.

What it loads

repo_id offers two variants: Qwen/Qwen3-TTS-12Hz-1.7B-Base (the default, the quality pick) and Qwen/Qwen3-TTS-12Hz-0.6B-Base (the fast/small pick). First run auto-downloads the weights into ComfyUI/models/Qwen3-TTS/ - and like the ASR loader, it first checks your existing HuggingFace/ModelScope caches and copies from there rather than re-downloading. A companion tokenizer repo (Qwen3-TTS-Tokenizer-12Hz) gets pulled in as part of the load.

The knobs are the same family as the ASR loader, so if you've used that node you already know these:

• source - HuggingFace or ModelScope, for when one is slow or blocked.
• precision - bf16 default (falls back to fp16 on Apple Silicon), fp16, or fp32.
• attention - auto default; flash_attention_2, sdpa, or eager if you want to push it.
• local_model_path - optional path under ComfyUI/models/ to load from disk instead of downloading.

The output and the cache design

Like every loader in this pack, the output is a model_key - a handle, not the model object. The actual weights live in an in-process cache; Voice Clone Prompt and SRT To Audio look the model up from that key. Consequences worth knowing:

• Change any loader setting and the cache is invalidated and the model reloads.
• The model sits in VRAM until you unload it. VoiceBridge Unload Model exists for exactly this, and you'll want it: the TTS 1.7B plus a still-resident ASR 1.7B plus your diffusion model is a crowded card on anything under 12GB. Load → generate → unload is the intended rhythm, and the unload counters feed back into the loaders so they know to reload cleanly.

Installing it

Part of the comfyui_voicebridge pack. ComfyUI Manager search "VoiceBridge", or:

cd ComfyUI/custom_nodes
git clone https://github.com/YanTianlong-01/comfyui_voicebridge.git
cd comfyui_voicebridge
pip install -r requirements.txt

Restart ComfyUI, then expect a multi-GB download on the first TTS load.

One honest warning

This pack is brand new - zero impressions, no community write-ups - so you're ahead of the wave either way. The model behind it is well-regarded (the community's complaints about the Qwen3-TTS line were mostly "give us the open weights," which is what these Base models are), but the pack's own integration is unproven at scale. If the TTS stage misbehaves, the [VoiceBridge] console logs are your first diagnostic, and remember the decoupled escape hatch: VoiceBridge SRT Splitter → any other TTS node → VoiceBridge Audio List Merger by SRT lets you swap in VoxCPM, Fish Audio, or whoever without touching this loader at all.