Qwen3 TTS Voice Design (Fast)

"Create a voice from a description" is the feature that made people stop and look at Qwen3-TTS when it shipped. You write something like "a calm, warm 40-something female audiobook narrator, clear but soft" and the model actually builds a voice that fits - no reference clip, no cloning, no consent questions. This node is that feature, exposed as: one text field for what to say, one for who should say it.

That makes it the fun one in this pack, and the one with a genuinely different workflow. Voice cloning answers "make this specific person speak." Voice design answers "invent a voice that matches this vibe," which is what you want for character work, for keeping narration voices consistent without owning the source audio, or for demoing ideas fast. The price: you're describing rather than demonstrating, so the output is a strong interpretation of your description, not a specific person. Adjust the description and re-run; it responds to detail.

How it works

VoiceDesign is a separate checkpoint, not a mode of the Base model. Load Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign in the Fast Loader - it's the only VoiceDesign option in the list, and the design module lives in those weights. This node then hands your voice_description to the model as the instruction that conditions the generated speaker, alongside the text you want spoken. The 24kHz mono AUDIO comes out the bottom.

Inputs you'll actually touch:

• voice_description - the whole game. Be concrete: age, gender, tone, pacing, accent, setting ("podcast host," "soft robot," "1930s radio announcer"). The richer the description, the less generic the voice.
• text - what gets spoken, multiline.
• language - Auto handles English fine; explicit options cover Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.
• seed - voices are sampled, so the same description with a new seed can drift. Found one you like? Keep the seed.

The usual suspects (max_new_tokens 2048, temperature 1.0, top_p 0.8, repetition_penalty 1.1) sit in optional. Same caveat as the rest of this pack: in the fast path, temperature only steers the talker stage, so it's seasoning, not a voice switch.

How to install it

Same as every node in the pack - ComfyUI Manager, search "Qwen3-TTS", install, restart. Manual:

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

NVIDIA GPU required. The VoiceDesign checkpoint downloads itself into ComfyUI/models/TTS/ on first run - the 1.7B download plus the graph-capture pass means the first run is the slow one, so don't panic when it sits there.

Common issues

Error right at the generate step - you loaded a CustomVoice or Base checkpoint instead of the VoiceDesign one. Check the Loader's model_name; the design module only exists in 1.7B-VoiceDesign.

Every voice sounds same-y - your description is too generic. "A female voice" gets you the platonic average; "a husky late-30s radio announcer who emphasizes every other word" gets you something you'd recognize. Also try changing the seed between attempts.

Long paragraphs get cut off / clamp warning - max_new_tokens hit the static-cache budget and got clamped to max_seq_len - 256. Raise max_seq_len on the Loader and re-run.

Worth knowing for a product context: unlike F5-TTS (CC-BY-NC) or Higgs v3 (non-commercial), the whole Qwen3-TTS stack here is Apache-2.0, so a designed voice doesn't quietly disqualify you from shipping. The license check on the TTS layer actually matters - this is one of the clean ones.