Breeze TTS 2 · T8 声音克隆

Voice cloning is the reason most people install a TTS pack, and T8_BreezeTTS_CloneRequest is the node that does it: feed it a reference recording plus an accurate transcript, and it builds a request that makes the model speak your text in that voice. It's zero-shot - no fine-tuning, no training run, no hours of data. The model's own description says it plainly: a clone request created from reference audio and its precise transcript.

The one input that decides everything

The tooltip on reference_text doesn't hedge: it must correspond exactly to the reference audio. This is where people burn their first hour. The model needs the verbatim words of the clip to align the encoded audio with the text tokens - if the transcript is wrong or just approximate, the alignment drifts and you get a warbly, half-robot clone that you'll chase through every other knob to fix. So:

• reference_text - type out exactly what's said in the clip, including hesitations if they're audible.
• reference_audio - a clean clip, and shorter than you think. The pack rejects anything over 60 seconds, checked on the raw waveform before it ever reaches the GPU or codec. Ten to fifteen seconds of clear, solo speech beats sixty seconds of a rambling recording every time.
• text - what you want the cloned voice to say.
• instruction (optional, default "Speak clearly and naturally.") - a natural-language performance direction layered on top of the cloned identity. You can keep the voice and still ask for energy, warmth, or urgency here.
• cfg_scale (optional, default 1.0) - interestingly, clone mode ships with CFG off at 1.0. The clone path is already strongly conditioned by the reference, so the pack's authors left guidance at neutral. Raise it if your clone sounds unsteady; leave it alone otherwise.

How the clone actually gets made

Breeze TTS 2 is an autoregressive model on a Qwen3 backbone, with a T5Gemma2 text encoder and a Qwen3-TTS codec turning audio into tokens. Cloning works by putting the transcript text and the encoded reference audio into the prompt as a "this is who's talking" segment, then appending your instruction and target text. The model conditions everything on those reference audio tokens, so it copies the timbre, pacing, and prosody - not just a generic voice.

One detail worth knowing: reference audio is cached. Encode the same clip once and it's keyed by a hash of the waveform, so re-running with the same reference doesn't re-encode it each time (the cache holds 8 clips, LRU-style). That's why iteration on text and instruction is cheap while the reference stays put.

Wiring and install

T8 模型加载器 → T8 声音克隆 → T8 生成设置 → T8 生成音频 → SaveAudio

The single output, request (type BREEZE_T8_REQUEST), feeds the Generate node. Install the pack via ComfyUI Manager (search "Breeze TTS 2 · T8star-Aix") or git clone https://github.com/T8mars/Comfyui-breeze-tts into ComfyUI/custom_nodes, then pip install -r requirements.txt. The model downloads automatically on the loader's first run once you've accepted the license.

The part the README won't advertise

Cloning is precisely the use the model license restricts. Breeze TTS 2 is research/non-commercial, and non-consensual voice cloning is explicitly prohibited. The node can't stop you, but the license can. Clone your own voice, voices you have permission to use, or public figures in clearly authorized contexts - and don't ship any of it commercially. For everything else, the Design node gives you a synthetic voice with none of the consent baggage, and the Direction node reuses a reference clip to change mood while keeping the identity.