Breeze TTS 2 · T8 声音设计

This is the node that makes Breeze TTS 2 feel like magic: you type a description of a voice, and the model invents a speaker to match. No reference audio, no transcript, no "whose voice am I allowed to clone" debate - the voice never existed before you asked for it. In the pack's three request modes, Design is the one with the lowest friction and the widest "I just need a narrator" use case.

What it's for

Say you need a warm, confident female narrator for a video voiceover, or a dry, mid-aged announcer for a product demo. Most open TTS forces you to either clone an existing clip or settle for a fixed set of voices. Breeze TTS 2's design path instead does what the model calls zero-sample voice design: the voice description is the speaker reference. It's the obvious first choice when you don't have a clean recording of anyone - and the honest answer to "where do I get the reference audio?" is that you don't need one.

The inputs that matter

• text - what gets spoken.
• voice_description - who's speaking. The author's tooltip is precise: describe the voice with no reference audio, and keep it in the same language as the text. A Chinese description driving English text is asking the model to thread a needle it shouldn't have to.
• cfg_scale (default 4.0, range 0.1–10) - guidance strength. This is the one knob to touch if the voice comes out flat or drifting: raise it toward 6–7 for a more committed delivery, drop it lower for a looser read.

The default description is a young, warm, confident female voice with clear diction - a perfectly usable narrator out of the box, which means you can wire this node up and hear audio before you've written a single line of your own description.

How it works under the hood

Design mode builds the prompt as a special instruction tag wrapping your voice description, followed by the text to speak. The model treats that description as the speaker condition and generates a consistent identity for the whole clip. Because there's no reference audio, there's also no reference encoding step - which makes Design the fastest of the three request modes and the easiest to iterate on: change a word in the description, re-run, hear the difference.

The other half is classifier-free guidance. The node builds a negative prompt from the same text without the voice description, and cfg_scale decides how hard the generation is pushed toward your description and away from the plain default. At the default 4.0 it's active; set it to 1.0 and it effectively disables CFG.

Wiring it up

The full chain, straight from the pack's example workflow:

T8 模型加载器 → T8 声音设计 → T8 生成设置 → T8 生成音频 → SaveAudio

In English: ModelLoader → DesignRequest → GenerationSettings → Generate → SaveAudio. The output is a single request of type BREEZE_T8_REQUEST, which feeds the request input on T8_BreezeTTS_Generate. Nothing else consumes it.

Gotchas worth knowing

Describe character, not effects. "A husky 50-year-old who speaks slowly and pauses a lot" will get you somewhere; "a voice with reverb" is asking the TTS to do audio engineering it doesn't do. And because Breeze TTS 2's license is research/non-commercial, this node's convenience doesn't extend to a paid product - it's for personal projects, dubs, and prototypes. For the many voices of one character shifting mood, you actually want the Direction node (keeps the identity, changes the take); Design is for voices that shouldn't exist yet.