FL Qwen3 TTS Voice Design

Most open TTS can only do one thing with a voice: clone one you already have. Qwen3-TTS Voice Design is the part that breaks the mold - you type "a warm, gentle female voice with a slight British accent" and the model builds a voice from the description. No sample clip, no recording, no cloning pipeline. It's the single feature in this pack that doesn't have an obvious rival in the usual open-TTS lineup, and it's why the Qwen3-TTS release got as much attention as it did back in January 2026.

The tradeoff is that you're trusting the model's interpretation of your words, so the first output is rarely the final one. The community's experience lines up with that: describing voice traits in plain language works, but it's iterative - you'll tweak the description more than you'll tweak the sampling settings. Start broad, lock in the voice, then refine.

How it works

Like the rest of the generate nodes, it takes a QWEN3TTS_MODEL from the Model Loader - and it insists on the VoiceDesign variant (Qwen3-TTS-12Hz-1.7B-VoiceDesign). Wire up the Base or CustomVoice model instead and you get the now-familiar "wrong model type" error telling you exactly what to load.

The voice_description field is the star. It's passed through as the model's instruction - the thing that controls timbre, accent, age, emotion, even persona. The default ("A warm, gentle female voice with a slight British accent") is a good template: mention the traits you actually care about. Keep it in the same spirit as prompting a diffusion model - concrete adjectives over vibes.

The inputs that matter

• voice_description - the voice recipe. Multiline. "A deep, gravelly male voice with a slow drawl" works better than "sounds cool."
• text - what the voice says. Multiline.
• language - the same ten languages as the rest of the pack. A voice designed in English will still speak Japanese; the description shapes the voice, the language shapes the speech. Describing traits in the language you're targeting can help the accent land.
• The usual sampling block - top_k, top_p, temperature, repetition_penalty, max_new_tokens, seed. Defaults are fine for 90% of use; temperature is the one you'd nudge up if output feels flat.

Output: one audio (AUDIO) connection, ready for preview or save.

Installing

Standard for this pack. ComfyUI Manager → search "FL Qwen3 TTS" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-Qwen3TTS.git
cd ComfyUI-FL-Qwen3TTS
pip install -r requirements.txt

Restart. First generation pulls the VoiceDesign model into ComfyUI/models/tts/Qwen3TTS/ - a few GB on first run.

A trick worth knowing

Voice design and voice cloning aren't rivals here - they chain. A trick people use: design a voice you like, generate a short sample with it, then feed that sample into the Voice Clone node as reference audio. You get the designed character with the extra control cloning gives you. For repeated multi-generation work, lock in the seed once you've found a design you like, because a designed voice is a sampled thing - same description, different seed, slightly different voice.