Aero-Ex/ComfyUI-MOSS-Diarize

MOSS-Transcribe-Diarize for ComfyUI

Single-file checkpoint of OpenMOSS-Team/MOSS-Transcribe-Diarize (0.9B, speech-to-text with speaker labels and timestamps) for the native ComfyUI node pack ComfyUI-MOSS-Diarize. It does not load in transformers.

File	Size	Notes
moss-transcribe-diarize-0.9B.safetensors	1.8 GB	bf16 weights, with config and tokenizer embedded.

Put it in ComfyUI/models/mossd/, then use Load MOSS-Diarize Model → MOSS Transcribe + Diarize.

Notes

• Same weights as upstream, repacked into one file. Greedy decoding was checked token-for-token against the upstream remote code in fp32.
• Each audio is transcribed in one pass, like upstream, so speaker labels stay consistent across the whole file.
• Outputs: a [start][Sxx]text[end] stream, segments JSON and SRT.

License

Apache-2.0, as the original release.