🎨 Voice Designer

Sometimes you don't have a reference clip and you don't want to comb through preset speakers - you just know the voice in your head. "A warm, confident older woman with a slight rasp." Voice Designer is the node that takes that sentence and makes the voice, no reference audio required. It's reference-free voice creation, and the output is a reusable voice you can save and drop into any TTS generation.

This node is a shared front-end over several voice-design-capable engines: Qwen VoiceDesign, MOSS VoiceGenerator, or OmniVoice in voice-design mode. You wire one of those in as the TTS_engine, write your description, and it produces a NARRATOR_VOICE plus a preview so you can hear it before committing.

How it works

The key design choice here - and it's a good one - is that the voice instruction lives on this node, not on the engine. The engine holds model/language/generation settings; Voice Designer holds the creative description. You give it a voice_instruction (what the voice should sound like) and a reference_text (the words it should speak in the preview), it generates, and out comes a voice object you can route straight into a TTS node or persist with the 💾 Save Character Voice node. Same seed + same instruction = the same voice, so once you find one you like, you can reproduce it.

The inputs and outputs that matter

• TTS_engine - required, and it must be a voice-design-capable engine (Qwen VoiceDesign, MOSS VoiceGenerator, or OmniVoice set to Voice Design). Plug in a plain cloning engine and it won't do voice design.
• voice_instruction - the description. This is the creative control: "A deep, authoritative male voice with clear articulation." Be specific about age, tone, pace, texture.
• reference_text - the sample line the preview speaks. The default is a friendly paragraph; swap in text that shows off the range you care about.
• seed - locks the voice. Change it to roll a different interpretation of the same description; keep it to reproduce a voice you liked.

Three outputs: opt_narrator (NARRATOR_VOICE) - the reusable voice, wire it into a TTS Text/SRT node or into 💾 Save Character Voice to keep it; preview_audio (AUDIO) - listen before you commit; and voice_info (text).

Installing it

Ships with the pack. ComfyUI Manager → search "TTS Audio Suite" → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py

No download for this node itself - the model behind whichever voice-design engine you connect (Qwen 1.7B VoiceDesign, MOSS, OmniVoice) is what downloads on first use.

Common issues & troubleshooting

"The instruction field is disabled / ignored." Voice design only works with the compatible modes. If you connected an engine whose current mode can't do voice design, the suite disables the instruction and, in mismatched cases, stops with a direct correction message rather than silently doing the wrong thing. Set the engine to a voice-design variant (Qwen VoiceDesign, MOSS VoiceGenerator, or OmniVoice Voice Design mode).

Every generation sounds different. You're rolling the seed. Voice design is stochastic - lock the seed once you find a voice you want to keep, and the same description will reproduce it. Then save it with Save Character Voice so you never have to re-roll.

I want to reuse this voice everywhere. Route opt_narrator into 💾 Save Character Voice. It writes the suite's three-file voice format under models/voices/, and from then on the voice is just another entry you can pick in the TTS nodes.

The voice is close but not quite. Tighten the voice_instruction - vague adjectives get vague voices. Naming concrete qualities (pitch, age, pace, accent, texture) steers it far better than "nice voice." The 📐 Visual Tag Builder can help assemble a structured description if freehand isn't landing.