MOSS-Transcribe-Diarize

FunASR integrates the third-party Apache-2.0 model OpenMOSS-Team/MOSS-Transcribe-Diarize published by OpenMOSS. It jointly generates long-form transcription, timestamps, and anonymous labels such as S01 and S02, without an external VAD or speaker pipeline.

Output contract

FieldMeaning
textReadable transcript without MOSS control tags.
timestampSegment-level start and end times in milliseconds.
sentence_infoNormalized segments with start, end, text, and anonymous spk.
raw_textExact tagged model generation retained for auditing.

FunASR OpenAI-compatible service

python -m pip install "transformers>=5.6,<6" fastapi uvicorn python-multipart
funasr-server --model moss-transcribe-diarize --device cuda:0 --port 8000

curl -fsS http://127.0.0.1:8000/v1/audio/transcriptions \
  -F file=@meeting.wav \
  -F model=moss-transcribe-diarize \
  -F response_format=verbose_json

The response preserves native anonymous speaker segments. A generic spk=true field does not start a second diarization model for MOSS.

Deployment paths

PathUse whenContract
FunASR serverYou need a self-hosted HF GPU endpoint, Docker Compose, or Kubernetes.verbose_json with speaker segments.
vLLMYou operate its scheduler and need the native audio endpoint.diarized_json; FunASR can normalize it through AutoModel.
SGLang OmniYou use its native MOSS pipeline.verbose_json; FunASR validates speaker prefixes before normalization.
LocalAI / moss-transcribe.cppYou need a third-party GGUF CPU or desktop-edge path.Independent C++ implementation; validate its output separately.
FunClipYou need speaker-aware SRT and clips.Consumes the normalized sentence_info contract.

Open the production deployment hub for complete pinned commands →

Verified service smoke

On one H100 80GB with Transformers 5.16.0.dev0, Torch 2.11.0+cu130, and model revision e8681d68, the real FunASR HTTP endpoint processed the bundled 6.0-second sample and returned HTTP 200, duration=6.0, and one non-empty monotonic segment labelled S01. This proves the service contract only, not accuracy, throughput, concurrency, or capacity.

Production boundaries

  • MOSS is owned and published by OpenMOSS; FunASR owns this adapter and deployment contract.
  • Pin and audit the model revision because the HF path uses trust_remote_code.
  • MOSS is an offline long-form model and is not exposed by the realtime WebSocket service.
  • Speaker labels are anonymous clustering labels, not known-person identity or voiceprint verification.
  • Do not externally VAD-split a meeting when global speaker continuity matters.

Repository guide · OpenMOSS upstream · FunClip integration