MOSS Whisper 词级对齐模型 · T8star-Aix

Here's a fact about MOSS that a lot of people discover the hard way: its native output is sentence- and segment-level timestamps. Great for subtitles at the cue level, useless if you need to know exactly when a word was said - for karaoke, for keyword search, for syncing a transcript to a video frame. This loader exists because the pack's answer to that is to bring in a second, separate model: a pinned openai/whisper-small, run purely to generate word anchors that get mapped back onto MOSS's transcript.

T8_MOSS_WordAlignmentModelLoader is the config-and-cache half of that arrangement. It doesn't do the alignment itself - that's T8_MOSS_WordAlignment, the node that actually runs Whisper over your audio and reports how well the word anchors matched your transcript. This loader just prepares the T8_MOSS_WORD_ALIGNMENT_MODEL handle. You wire it, you set maybe two things, and it quietly does its job.

The settings that matter:

• model_id - defaults to openai/whisper-small on Hugging Face, pinned to revision 973afd24965f72e36ca33b3055d56a652f456b4d. The pack deliberately fixes the revision so your alignment is reproducible; you can also point this at a local model directory.
• language - default auto. If you know the content language, set it (zh, en, …). This is genuinely worth doing: Whisper does faster and more accurate word-level alignment when it doesn't have to guess the language first, and it's the one input beginners usually ignore.
• chunk_length_seconds - 30 by default (5–120 range). Longer chunks mean fewer passes but more VRAM per pass; drop it if you're on a tight card.
• device and precision - both auto, same resolution logic as the rest of the pack.
• release_after_run - true by default, so the aligner frees its VRAM after the alignment node finishes.

It outputs the aligner handle plus a JSON info string describing what got configured. The model only loads when the alignment node actually runs - lazy loading, independent cache, and - per the pack's design - this loader never touches ComfyUI's own torch or transformers stack, so it can't break your image models.

A note on expectations, because this is where people get let down: Whisper-small is a tiny model and word-alignment coverage is not 100%. In the pack author's own benchmark runs against real speech, model-matched coverage landed between roughly 55% and 72%. That's normal and by design - the alignment node marks unmatched words as interpolated instead of pretending they came from the model, and it flags coverage below 50% as a warning. So don't panic when the coverage report isn't perfect; do read it.

Installing the pack

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-MOSS-Transcribe-Diarize-T8.git comfyui-MOSS-Transcribe-Diarize-T8

In the ComfyUI Python environment (..\..\python_embeded\python.exe on Windows Portable): pip install -r requirements.txt, run scripts/check_transformers.py (needs Transformers >= 5.5.0, < 6), restart ComfyUI. ComfyUI Manager users: search comfyui-moss-transcribe-diarize-t8. The Whisper-small weights download automatically on this loader's first execution - no manual model download needed for the aligner.

One last thing: word alignment only makes sense if you actually feed it the same audio and transcript the main model produced. Wire it into the standard chain - main model → transcript → this loader → T8_MOSS_WordAlignment → subtitle export - and you'll get cues with per-word anchors for free. Try to align against a transcript from a different source and the coverage report will lovingly tell you how badly it mismatched.