Breeze TTS 2 · T8 桌面音色包

If you've tuned voices in T8star-Aix's desktop Voice Studio app, you've probably exported a .t8voice.zip - a self-contained voice pack carrying the speaker's reference audio, a saved delivery instruction, and a default CFG. This node ("T8 桌面音色包", desktop voice bundle) is the bridge that gets those voices into ComfyUI. Point it at a bundle file, type a line, and out comes the same request the Design and Clone nodes produce - meaning the voice you already auditioned in the desktop app works in a node workflow with zero re-tuning.

What it actually does

Give it a local path and it reads the bundle offline and without ever extracting to disk - every member is parsed in memory. That "offline, no extraction" framing isn't paranoia fluff: the source treats a .t8voice.zip as untrusted input and runs a real gauntlet of checks before decoding anything. Rejected outright: absolute or path-traversal member names, symlink bundle paths, Windows-unsafe duplicate casing, encrypted members, undeclared members, and absurd compression ratios. Everything is capped - member count, archive and decompressed size, reference-audio duration (60 seconds max, same as the rest of the pack) - and declared sizes are verified against SHA-256 hashes. If you're in the habit of downloading random voice packs, this is the good kind of gatekeeping.

Once the bundle passes, the embedded profile provides the voice's id/name, language, stored instruction, default cfg_scale, and the reference transcript, and the reference audio is decoded straight to standard ComfyUI AUDIO format ([batch, channels, samples] float32). No round-trip through your disk or a temp file.

The inputs that matter

• bundle_path - absolute path to the local .t8voice.zip. The default "voice.t8voice.zip" is a placeholder; the shipped voice_bundle_workflow.json example needs you to swap in a real path too.
• text - the line to synthesize. Inline vocal events go right in here, unmarked and unstripped: Chinese [笑] [咳嗽] [清嗓子] [叹气], English (laugh) (cough) (clears throat) (sigh).
• line_direction_mode - inherit uses the bundle's saved voice instruction; override replaces it with your line_direction for this line; neutral applies a clean, natural delivery. As with the pack's per-line node, override requires line_direction or it errors out.
• cfg_scale - 0 (default) inherits the bundle's stored CFG; anything else overrides it for this line.

The three outputs

This is where the node earns its keep, because it returns three things:

• request - a BREEZE_T8_REQUEST; wire it into T8_BreezeTTS_Generate.
• reference_audio - the standard ComfyUI AUDIO. Preview it with PreviewAudio or SaveAudio to hear the voice you're about to use, no synthesis required. You can also feed it to other audio nodes.
• voice_info - a JSON string describing the voice profile and the effective mode the bundle resolved to.

That middle output is quietly brilliant for voice-library browsing: audition the actual reference before spending a generation on a take.

Installing it

Search "Breeze TTS 2 · T8star-Aix" in ComfyUI Manager and install, then restart. Manually, the same thing looks like:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-breeze-tts.git comfyui-breeze-tts-T8
python -m pip install -r comfyui-breeze-tts-T8/requirements.txt

The pack only pulls lightweight deps (soundfile, safetensors, huggingface_hub, …) - torch/transformers belong to your ComfyUI host and are never touched. It does demand transformers >=4.57,<6 and Python 3.10–3.12, and reports at registration if your host falls outside that. First run of the Model Loader downloads the actual BreezeBlue/Breeze-TTS-2 weights (pinned revision) into ComfyUI/models/breeze_tts/, gated behind the accept_model_license checkbox - Breeze TTS 2 is research/non-commercial, not open source.

Troubleshooting

Three things bite people. Wrong path: the node wants an absolute path to a real .t8voice.zip, not a symlink - "please choose a desktop-exported .t8voice.zip" means it didn't find one. Decode failures: if the bundled reference won't decode, re-export it from the desktop app as WAV or FLAC, which the README calls out as the most stable cross-platform choice. Schema mismatch: the node tracks the desktop app's release cadence, so if it rejects a freshly exported bundle, update the desktop app first. And remember - the bundle is read in memory, so keep the file where you pointed it; the node re-fingerprints the file each run and picks up changes without a restart.