Qwen3 TTS Custom Voice (Fast)

This is the node to reach for when you want decent narration right now and don't want to hunt down a reference clip to clone from. Qwen3-TTS ships a set of preset voices - Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee - and this node speaks your text in whichever one you pick. No reference audio, no transcript, no setup. Type, pick a voice, run.

It's the fastest path from "I need a voiceover for this clip" to audio inside the whole pack. If you've ever burned an afternoon trying to clone a voice that then drifts on the second paragraph, you'll appreciate that these presets are stable - same voice every run, which is exactly what you want when the voice is incidental (a test narration, a talking-head placeholder, a quick audio pass before you commit to a real clone).

How it works

The "CustomVoice" in the name is the model family, not this node's job. You load a CustomVoice checkpoint (0.6B or 1.7B) in the Fast Loader - those checkpoints carry the speaker embeddings baked in - then this node calls the model's custom-voice generation with your text and speaker. The 24kHz mono AUDIO it returns plugs straight into a preview/save node.

The inputs you'll actually set: text (multiline), speaker (the nine presets), language (Auto is fine for English; the list covers Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian), and seed if you want reproducible takes. The optional instruct field is worth knowing about - it's a plain-language instruction for how to deliver the line (calmer, more excited, whisper) rather than what voice to use. Left empty, it's skipped entirely.

The sampling trio (temperature 1.0, top_p 0.8, repetition_penalty 1.1) and max_new_tokens (2048) sit in optional land. One honest caveat from the pack's own docs: in this fast path, temperature and top_p only steer the talker stage - the code predictor's sampling is fixed at CUDA-graph capture time, so don't expect temperature to give you wildly different deliveries. It's for nudges, not character switches.

How to install it

Same install as the rest of the pack - ComfyUI Manager, search "Qwen3-TTS", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/YildirimMC/ComfyUI-Qwen3-TTS-Fast
pip install -r ComfyUI-Qwen3-TTS-Fast/requirements.txt

Needs an NVIDIA GPU. The first time you run, the CustomVoice model downloads itself into ComfyUI/models/TTS/ - no manual model hunting. Note it needs its own loader run: you can't grab a Base model (for cloning) and feed it to this node expecting built-in speakers to appear.

Common issues

Nothing happens / instant error - you wired up a Base or VoiceDesign checkpoint, not a CustomVoice one. Go back to the Loader and switch model_name.

It's slow on first run - that's the model download plus the CUDA graph capture pass, which happens once per (model, device, dtype, attention, max_seq_len) combo. Subsequent runs are fast.

Long text gets truncated or you see a clamp warning in the console - max_new_tokens exceeded the static-cache budget (max_seq_len - 256), so it got clamped. Raise max_seq_len on the Loader if you're generating paragraphs, and remember it needs a reload after you change it.

If you're here to build a talking-head pipeline, the common shape is: this node makes the voice → an audio-driven avatar model (think the InfiniteTalk/MultiTalk corner of the ecosystem) makes the face move. Wire the AUDIO output into whichever node in that pack eats audio.