Qwen3-TTS Voice Design

Qwen3-TTS Voice Design is the pack's party trick: instead of picking a preset voice or cloning one, you describe the voice you want in plain English and it invents one. "A deep, resonant male voice, narrator style, calm and professional" - that's the whole input. It's the text-to-voice equivalent of prompt-based generation, and it's the thing that separates Qwen3-TTS from most of the open TTS scene, which still makes you choose between cloning a voice or settling for a preset list.

This is the model's headline feature (Alibaba markets it as "from what to say to how to say it"), and in ComfyUI this node is where you actually touch it. It pairs with the VoiceDesign model variant from the Qwen3-TTS Loader - Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign is the only option, since there's no 0.6B VoiceDesign. Wire a QWEN3_MODEL in, type your text and your instruct voice description, and you get an AUDIO out. language supports the same ten-language list as the rest of the pack (or Auto), and seed defaults to 42 like everywhere else in this suite.

How it works

The VoiceDesign variant is a distinct model checkpoint whose instruction-following head maps your natural-language description onto the model's timbre/prosody/emotion latent space, then synthesizes the text in that voice via the same autoregressive pipeline as the other generators. In practice that means it's not a search over a fixed voice library - the description genuinely conditions generation. That's also why it needs its own weights and why you get a hard "does not support generate_voice_design" error if you feed it a CustomVoice or Base model.

Honest expectations, because the community has tested this one hard: the descriptions steer the voice noticeably - gender, register, pace, narration-vs-conversational all land reasonably well. Where it struggles is fine detail. People report style descriptions getting flattened toward a generic American-accented read, and it's not going to nail "slightly hoarse, like a detective who's been yelling all night." Treat it as a directional dial: it gets you in the neighborhood of a voice, and you iterate on the wording of instruct the way you'd iterate on an image prompt. Write concrete adjectives, not vibes.

Using it well

The instruct field is the one that matters, and it rewards specificity. Compare "male voice" (boring, generic) to "a low, gravelly, authoritative male narrator, deliberately paced, professional documentary tone" (actually different). You can stack emotion and delivery hints into the same description. If you want a voice that persists across many lines, you're better off generating once, saving it, and driving future runs off a prompt embedding - but Voice Design doesn't expose that directly, so for multi-sentence work you'd clone from the result instead.

Installing & troubleshooting

Same pack install as all the Qwen3-TTS nodes: ComfyUI Manager (search "ComfyUI-Qwen3-TTS") or git clone https://github.com/DarioFT/ComfyUI-Qwen3-TTS into custom_nodes, then pip install -r requirements.txt - ComfyUI won't do that step for you, and qwen-tts pinning transformers==4.57.3 can downgrade your environment. The 1.7B VoiceDesign model downloads on first use.

The two failure modes you'll actually see: the model-type error above (you loaded the wrong variant), and the standard Qwen3-TTS generation hang (GPU pegged, nothing completing). For the hang, lower max_new_tokens on the loader/generator path, nudge the seed, and restart ComfyUI if it's stuck. And if a description comes out sounding identical to the last one, that's the known flattening behavior, not a bug - rewrite instruct with more concrete voice adjectives and try again.