APZmedia: Voice Design Generate

The fastest way to make Qwen3-TTS talk: type some text, type how you want it said, hit run. APZmedia: Voice Design Generate is the single-call version of the pack's voice-design magic - no reference clip, no saved identity, no two-model dance. Just the VoiceDesign model, a line of text, and a natural-language direction, and out comes a AUDIO you can preview or save.

This is the node to reach for when you're iterating - testing voices, doing a one-off narration, or just finding out whether this whole TTS-in-ComfyUI thing is for you. Its sibling Voice Design exists to create a reusable identity; this one throws the voice away after each run. That's a feature, not a bug: it keeps the workflow dead simple and frees you from curating a library before you've figured out what you actually want to sound like.

The inputs

• text - what to say. Multiline, so paste whole paragraphs.
• instruct - the voice and delivery. Default "Speak clearly with a warm, friendly tone," but this is the field that separates good output from flat output. Try "Deep authoritative male, slightly amused" or "Hushed and conspiratorial, rushing slightly." The model is an instruct-following TTS; the more you direct, the more it performs.
• language - English default, ten more languages, or Auto.
• seed - set it if you want reproducible takes; leave it if you're hunting for a good one.

The model socket needs a VoiceDesign model from APZmedia: Model Loader (Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign). Hook a PreviewAudio node or the pack's Audio Saver onto the audio output and you're done.

A quick workflow

[Model Loader: VoiceDesign] → [Voice Design Generate]
                                   text: "Hello world"
                                   instruct: "Warm, friendly female"
                                        ↓
                                   [PreviewAudio] / [Audio Saver]

Gotchas

The output is a standard ComfyUI AUDIO dict - waveform tensor [1, 1, samples], float32 in [-1, 1], sample rate 24000 Hz - so it plugs into any native audio node, not just this pack's saver. Good to know before you reach for a converter.

If speech comes out monotone or rushed (a common first impression with Qwen3-TTS, and people do complain about it when comparing against VibeVoice), resist blaming the model and look at instruct. The community's single biggest mistake with this model family is feeding it no direction at all and expecting it to intuit emotion from text like other TTS engines do. Qwen3-TTS expects to be told. Empty or vague instructions get you default, flat delivery.

Install is the whole pack: clone https://github.com/APZmedia/ComfyUI-APZmedia-qwen-TTS into ComfyUI/custom_nodes, restart, and the first generation downloads the 1.7B model from Hugging Face automatically - that first run is slow, every run after is not.