FireRedAudio 声音设计 · T8star-Aix

Here's the version of voice cloning where nobody has to be cloned. FireRedAudio 声音设计 takes a plain-language description of a voice - "warm, mid-forties, slight husk, unhurried narrator" - and synthesizes speech that sounds like that, with no reference audio at all.

That's a genuinely different job from the pack's TTS node, and it's the one you reach for when:

• You don't have a reference clip for the voice you want.
• You're inventing a character, and no real person should be the source.
• You want the safety (and legal clean-room) of a voice that belongs to nobody.

It's the same FireRedAudio model under the hood, just told what voice to make in words instead of being shown one.

How it works

The flow is the same machinery as the zero-shot TTS: the Qwen audio backbone generates speech tokens conditioned on your instruction (the voice description), and the RedAE flow decoder turns them into a 24 kHz WAV. Because it is generation, this node needs the full model profile - main model plus the RedAE decoder - so unlike the audio-understanding node, the lite download won't cover it.

The inputs that matter

• model - from the "FireRedAudio 模型/隔离运行时" loader.
• instruction - the voice description. This is the whole game. Be concrete: age, gender, register, energy, accent hints, pace. Vague prompts give you the model's default announcer voice, which is fine but boring.
• text - what the voice says.
• settings (optional) - the shared generation parameters node, where seed, quality preset, diffusion steps and CFG live. Same one the TTS node uses.

Outputs: 生成音频 (AUDIO - wire into the pack's "保存音频" node or any audio sink) and a 运行报告 with the generation stats.

Installing it

Standard two-step for the comfyui-fireredaudio-T8 pack:

cd ComfyUI\custom_nodes
git clone https://github.com/T8mars/comfyui-fireredaudio-T8.git
cd comfyui-fireredaudio-T8
python scripts\setup_runtime.py

The setup script builds the pack's isolated Python 3.10 / Transformers 5.8 worker so FireRedAudio's dependencies never touch your ComfyUI environment - don't skip it, the node will just error. Then the model, and this time the profile matters:

python scripts\download_models.py --target "D:\ComfyUI\models\TTS\FireRedAudio" --profile full --variant int8-wo-safe-v1

--profile full is required here. Voice design generates audio, and generation needs the RedAE decoder that only the full profile includes.

Where people get burned

First, set expectations: a described voice is less faithful than a cloned one. Cloning pins a real person's timbre; describing gives you "in the ballpark of." Plan on a few passes with the instruction before a character clicks - treat the description as a dial you're tuning, not a contract. Second, cold start pays the same minute-or-two model-load tax as every generation node in this pack, and generation still spikes around 20 GB of VRAM with sequential offload auto-selected under ~36 GB free. Third, keep the instruction in plain terms - the model reads English and Chinese, but it's not a prompt-engineering riddle, and over-wrought descriptions read as noise. The one-liner I actually reach for: "warm, calm, mid-range male voice, clear diction, moderate pace." Works every time.