Voice Design (QwenTTS)

Preset speakers bore you, and you don't have a reference clip of the voice you want. That's the exact gap Voice Design (QwenTTS) fills: instead of picking a timbre or cloning one, you describe one. "A warm, gentle female voice." "A gravelly old sea captain." "An excited teenage sports announcer." The node takes that natural-language description plus your text and generates speech in a synthesized voice that matches it - no reference audio anywhere.

It's the feature that made Qwen3-TTS's launch announcement interesting in the first place. Voice cloning was old news by then; designing a voice from a description was the differentiator, the thing that frees you from both preset lists and clip-hunting.

How it works

This node loads the Qwen3-TTS-12Hz-1.7B-VoiceDesign model - note the size: there is no 0.6B VoiceDesign, so model_size is locked to 1.7B. That model accepts a free-text instruct describing timbre, prosody, emotion, persona, and more, and conditions generation on it rather than on a speaker embedding. The model also handles the language (Auto + 10 languages), and it's genuinely good at the instruction-following part - within limits.

The inputs

• text - what the voice should say.
• instruct - the voice description. This is the whole point of the node. Be specific: pitch, speed, age, gender, accent, mood. One line like the default ("A warm, gentle female voice.") works; five lines of detail work better.
• model_size - 1.7B only, as above.
• language - Auto or one of the ten. The README's tip applies harder here than anywhere: write and describe in the target language. The classic "my designed voice ignores everything and comes out American-accented" report is usually a language mismatch - English instructions on Chinese text, or Auto language guessing wrong.

Optional: seed (reproducibility), unload_models (default true). Output is a single audio (AUDIO) for your save/preview nodes.

Install

Standard for this pack - ComfyUI Manager (search ComfyUI-QwenTTS) or:

cd ComfyUI/custom_nodes && git clone https://github.com/1038lab/ComfyUI-QwenTTS.git
python3 -m pip install -r ComfyUI/custom_nodes/ComfyUI-QwenTTS/requirements.txt

The VoiceDesign 1.7B model auto-downloads to ComfyUI/models/TTS/Qwen3-TTS/ on first use (a few GB).

The honest caveats

Voice Design's output is the most hit-or-miss of the three Qwen3-TTS modes. The model is really good at producing a pleasant, expressive voice for broad descriptions and mediocre at precise, controllable accents. The community take, straight from a help thread: "Voice Design flat out ignores all styles and only outputs an American accent" - which, again, is usually a language/instruction issue, but it's also just what the model does at the limits. If you need a specific accent or a specific person, clone instead. If you need "some voice that feels like this," design. And if a description works out, Create Voice (QwenTTS) can't save it as-is (that node saves clones), so design is a one-shot-generate tool - keep the seed around if you want to reproduce a take.