Breeze TTS 2 · T8 声音导演

This is the node that sets Breeze TTS 2 apart from most open TTS, and it's the least obvious of the three request modes. T8_BreezeTTS_DirectionRequest keeps the reference speaker's identity but lets you direct their performance - emotion, pacing, delivery - via a plain-language instruction. Think of it as the difference between a voice double and a voice director: you're not swapping who's talking, you're changing how they deliver the line.

Where it fits

The other two request nodes are the obvious endpoints: Design invents a voice from a description, Clone copies one from a recording. Direction sits in the middle and does the thing both of them can't. With a clone you get the voice but roughly one delivery; with design you get a voice that never existed. Direction is for when you have a character whose voice you already like and you need them to say the same kinds of lines in different moods - a narrator who turns conspiratorial, a character who goes from calm to strained across a scene. That "keep the actor, change the take" workflow is real dialogue work, and this is the node built for it.

The inputs that matter

• text - the line to speak.
• reference_audio and reference_text - the same pairing as clone mode, with the same hard rule: the transcript must match the clip exactly. And the same 60-second ceiling on the reference, enforced before it reaches the codec. You cannot get the identity right if the transcript is wrong, so do this part carefully.
• direction - the whole point. The tooltip describes it as emotion, rhythm, and expression commands while the timbre stays unchanged. The default - "slow down, restrained and serious" - is a decent template for how to write these: short, behavioral, in the same language as the text.
• cfg_scale (default 4.0) - note the difference from clone mode. Direction ships with guidance on by default, because the instruction is the whole trick: the model builds a positive prompt (transcript + reference audio + direction + text) against a negative that has the direction stripped out, and CFG pushes the delivery hard toward what you asked. If the performance comes out muted, this is the first knob to raise.

How it works

Mechanically it's the clone path with a promotion: the reference audio gets encoded to Qwen3-TTS codec tokens, the transcript anchors them, and the direction text fills the instruction slot that clone mode leaves as "Speak clearly and naturally." Because the reference handling is shared, Direction gets the same audio caching as clone - same clip, no re-encode, cheap to iterate on the direction line until the take lands.

Wiring it up

T8 模型加载器 → T8 声音导演 → T8 生成设置 → T8 生成音频 → SaveAudio

ModelLoader → DirectionRequest → GenerationSettings → Generate → SaveAudio. The node outputs a single request of type BREEZE_T8_REQUEST for the Generate node - same contract as the other two request modes, so you can swap Design/Clone/Direction in and out of the same workflow without rewiring anything.

Practical notes

Write the direction like you'd write stage notes, not a wishlist: "faster, clipped, impatient" beats "convey a sense of urgency." And keep the reference transcript honest - a lazy transcript here doesn't just sound off, it degrades the identity you were trying to preserve. One caveat from the license file that applies to everything in this pack: Breeze TTS 2 is research/non-commercial, and the model's prohibition on non-consensual cloning extends to whatever identity you're directing. Directing your own voice, a voice you own, or a clearly authorized character is exactly what this node is for.