MOSS 提示词与热词 · T8star-Aix

The MOSS transcribe nodes don't just take audio - they take a prompt, because MOSS-Transcribe-Diarize is a generative audio model that produces text one token at a time. T8_MOSS_PromptHotwords is how you shape that generation. It builds a strict, formatted instruction that tells the model exactly what you want back: segments starting with a timestamp and a speaker number ([S01], [S02], …), the spoken content, and an end timestamp. The transcription nodes then parse that format into structured segments, SRT and ASS.

Think of it as the "make it useful" node. Without a solid prompt, the model will still produce text, but you'd be hand-parsing it. With it, you get clean, timestamped, speaker-tagged output the whole rest of the pack can consume.

The inputs you'll actually touch

• base_prompt - the editable instruction, pre-filled with the strict format described above. Leave it alone until you know what you're doing; it's the contract the parser depends on.
• hotwords - comma-separated proper nouns and names, default OpenMOSS, ComfyUI, T8star-Aix. Here's the honest part: hotwords are hints that get folded into the prompt, not enforced tokens. The README says so outright ("热词只作为提示，不保证强制命中"). If a name is getting mangled, listing it here nudges the model in the right direction - but it's not a guarantee the way Whisper-style forced hotwords are. Set it for the words that matter and temper your expectations.
• language_hint - auto, or one of 中文 / English / 日本語 / 한국어 / 粤语. A hint, not a lock. For genuinely mixed-language audio, auto is the safer pick.
• preset_id - a scene-preset dropdown; default is the one you'll use most.
• strict_format - on by default. Turn it off only if you want freeform output and don't mind losing the structured parsing downstream.

Outputs: MOSS 提示配置 (the typed prompt object you wire into T8_MOSS_TranscribeDiarize or T8_MOSS_SmartLongAudio) and 最终提示词 (the fully assembled text). The second one is great for a ShowText node - preview what the model is actually being told before you commit.

Installing it

It ships inside the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-MOSS-Transcribe-Diarize-T8.git comfyui-MOSS-Transcribe-Diarize-T8
cd comfyui-MOSS-Transcribe-Diarize-T8
pip install -r requirements.txt
python scripts/check_transformers.py      # requires Transformers >= 5.5.0
python scripts/download_models.py --comfyui-root ..\..

Restart ComfyUI and it shows up under the T8star-Aix category. The model itself is a ~1.8GB download into ComfyUI/models/moss_transcribe_diarize/, and this pack wants an NVIDIA card with 12GB+ for the "official" tier (8–10GB only for short clips).

Gotchas

• Don't read hotwords as a fix. If the model genuinely can't transcribe a word, the hotword list will help marginally, not magically. It's guidance in the prompt, not a token constraint. The author is refreshingly upfront about this.
• Don't vandalize base_prompt. The parser downstream expects [start][Sxx]text[end] structure. If you rewrite the prompt into prose and flip strict_format off, you'll get valid-looking transcription that the validate/quality nodes will rightly flag as unusable.
• The UI is language-aware. This pack follows ComfyUI's language setting, so field labels switch between Chinese and English with the rest of the interface. If your ComfyUI is in English you'll see English labels - don't be confused when a screenshot shows the Chinese defaults.