MOSS-Transcribe-Diarize joins the FunASR ecosystem / MOSS 一体化转写与说话人分段已接入 FunASR

[LauraGPT]

FunASR now provides a production-oriented integration for the third-party OpenMOSS MOSS-Transcribe-Diarize model. It performs long-form transcription, timestamps, and anonymous speaker labeling (S01, S02, ...) in one generation, so applications do not need to assemble a separate external VAD and speaker-diarization pipeline.

FunASR 现已接入第三方 OpenMOSS MOSS-Transcribe-Diarize 模型。它在一次生成中完成长音频转写、时间戳和匿名说话人标签（S01、S02……），业务侧不需要再组合外部 VAD 与说话人分离流水线。

Available paths / 可用路径

• FunASR AutoModel: local Transformers, an existing vLLM server, or the SGLang-compatible transcription contract.
• vLLM 0.27.1: OpenAI-compatible /v1/audio/transcriptions with response_format=diarized_json.
• FunClip 2.2.0: speaker-aware SRT and speaker clip export using MOSS segments.
• funasr.com: bilingual, revision-pinned deployment commands and validation boundaries.

Quick start with vLLM / vLLM 快速开始

Start the pinned model with a vLLM build that includes MOSS support:

vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize \
  --revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8 \
  --served-model-name moss-transcribe-diarize \
  --trust-remote-code \
  --port 8898

Use the public FunASR package to normalize the response into the common sentence_info contract:

python -m pip install -U "funasr==1.4.10"

from funasr import AutoModel

model = AutoModel(
    model="OpenMOSS-Team/MOSS-Transcribe-Diarize",
    backend="vllm",
    vllm_base_url="http://127.0.0.1:8898/v1",
    vllm_model="moss-transcribe-diarize",
    vllm_response_format="diarized_json",
    disable_update=True,
)

result = model.generate("meeting.wav", max_completion_tokens=8192)[0]
for segment in result["sentence_info"]:
    print(segment["start"], segment["end"], segment["spk"], segment["text"])

For long meetings, size max_completion_tokens against the longest expected recording. Do not attach vad_model or spk_model: external segmentation can break global speaker-label consistency.

长会议需要按最长录音评估 max_completion_tokens。不要再传入 vad_model 或 spk_model；外部分段可能破坏全局说话人标签的一致性。

Evidence and boundaries / 验证与边界

The pinned vLLM path was exercised on an H100 with a two-speaker probe and real Chinese long recordings. FunASR preserves speaker segments and timestamps in sentence_info; FunClip 2.2.0 consumes the same segments for SRT and per-speaker clipping.

MOSS-Transcribe-Diarize remains an OpenMOSS model under its upstream Apache-2.0 license. FunASR owns only the adapter, normalized result contract, tests, and deployment documentation. S01 / S02 are anonymous diarization labels, not enrolled-person identity verification. This is currently a long-form/offline path, not a replacement for Fun-ASR-Nano's low-latency realtime WebSocket service.

MOSS-Transcribe-Diarize 仍是 OpenMOSS 所有并维护的第三方模型。FunASR 提供的是适配器、统一结果契约、测试和部署文档。S01 / S02 是匿名说话人分段标签，不是已注册人员身份验证；当前路径面向长音频/离线处理，不替代 Fun-ASR-Nano 的低延迟实时 WebSocket 服务。

Links / 入口

• English deployment guide
• 中文部署指南
• Live English product page
• 中文产品页
• FunClip 2.2.0
• FunASR 1.4.10

Please share the model revision, runtime version, audio duration/languages, GPU, response format, and a redacted output excerpt when reporting results. We will keep reports open through reporter verification rather than treating a release as proof that every workload is resolved.

反馈时请附模型 revision、runtime 版本、音频时长/语言、GPU、response format 和脱敏后的输出片段。相关问题会保持开放等待报告者复测，不会把“已发布”直接等同于“所有场景已解决”。