docs: add FunASR ecosystem deployment path

[LauraGPT]

Summary

• add bilingual FunASR ecosystem entries to the Quickstart table of contents
• link the released FunASR 1.4.12 deployment path for vLLM, SGLang Omni, Transformers, and LocalAI/GGUF
• include a runnable AutoModel example for vLLM diarized_json speaker segments
• link FunClip 2.2.1 for speaker-aware SRT and per-speaker clip export
• clarify that applications do not attach an external VAD or speaker model, while OpenMOSS retains model ownership and Apache-2.0 attribution

Verification

• exact signed+DCO head: 43a7c31
• all Python fences in both READMEs parse
• PyPI FunASR 1.4.12, FunClip 2.2.1, and both deployed FunASR product pages return HTTP 200
• git diff --check passes

This PR changes README files only. The runtime and model implementation remain owned and maintained by OpenMOSS; FunASR provides the ecosystem adapter and deployment contract.

[LauraGPT]

The FunASR integration referenced by this documentation PR is now available in the public funasr==1.4.7 package:

• PyPI: https://pypi.org/project/funasr/1.4.7/
• GitHub Release: https://github.com/modelscope/FunASR/releases/tag/v1.4.7
• FunASR guide: https://github.com/modelscope/FunASR/blob/main/docs/moss_transcribe_diarize.md

The published wheel was rebuilt from exact tag commit 00dd1a54e077d0823263dfc4f6f01ef686b8fde9. We verified the PyPI and GitHub wheel/sdist bytes and SHA-256 values match, then installed from public PyPI and exercised the backend="vllm" client construction plus transcript/timestamp/anonymous S01/S02 normalization. The release also keeps the ownership boundary explicit: MOSS-Transcribe-Diarize remains an OpenMOSS third-party model; FunASR provides the adapter and deployment contract.

[LauraGPT]

The linked FunASR ecosystem path now includes the official vLLM structured speaker-segment contract end to end:

• merged FunASR adapter/docs PR: feat(moss): support vLLM diarized responses modelscope/FunASR#3566
• exact merge commit: 4300a6b2a158e126ff5d97ed55816b43ae88522a
• live English deployment page: https://www.funasr.com/en/deploy/moss-transcribe-diarize.html
• live Chinese deployment page: https://www.funasr.com/deploy/moss-transcribe-diarize.html

The opt-in FunASR setting vllm_response_format="diarized_json" maps vLLM 0.27.1 speaker segments directly to sentence_info; the existing json default remains compatible with raw tagged text. We verified both direct HTTP and the FunASR adapter on one H100 with the pinned model revision, and the PR exact head passed all three repository CI checks. The guide also records that the older 68b4a1d5 nightly predates vLLM #48543 and returns HTTP 400 for diarized_json, so operators do not silently retry the wrong contract. Ownership remains unchanged: OpenMOSS supplies the Apache-2.0 model and FunASR supplies the ecosystem adapter and deployment contract.

[LauraGPT]

Updated the exact signed+DCO head to a15e215 with a runnable bilingual FunASR vLLM example.

The README path now starts from the public funasr>=1.4.8 package, requests vLLM diarized_json, and shows the normalized sentence_info speaker segments. It still states that no external VAD/speaker model should be attached and keeps OpenMOSS ownership explicit.

Fresh validation:

• all Python fences in both READMEs parse
• FunASR 1.4.8 MOSS model/docs suite: 17/17 passed
• PyPI, FunASR guides, and both deployment pages return HTTP 200

[LauraGPT]

@zimoCCC could you review this docs-only ecosystem entry when convenient? The PR is mergeable and now points to the public funasr>=1.4.8 AutoModel/vLLM path. Its exact signed+DCO head a15e2156cd20c54eceecf00dfe0ddb140c2bcb78 passed 17/17 FunASR MOSS tests, and both README Python examples parse. I cannot submit a formal reviewer request from the fork, so this is the one maintainer mention; I will avoid repeated pings.

[LauraGPT]

Substantive downstream update: FunASR PR modelscope/FunASR#3579 has now merged as d5a8db7fef9c0a85c734f30da9bed1d413c734f8.

FunASR now exposes MOSS-Transcribe-Diarize through AutoModel on HF, vLLM, and SGLang paths, while preserving speaker-attributed segments and timestamps and clearly retaining OpenMOSS ownership/provenance. The exact-head adapter, documentation, product-site, and browser checks passed. The bilingual deployment pages are live at:

• https://www.funasr.com/deploy/moss-transcribe-diarize.html
• https://www.funasr.com/en/deploy/moss-transcribe-diarize.html

Boundary: the SGLang path is contract-tested against the OpenAI-compatible transcription response and pinned to the upstream revision that forwards max_new_tokens; we are not claiming a new live-H100 benchmark from the FunASR host because its current CUDA/driver boundary does not match that pinned SGLang build. This PR remains useful as the OpenMOSS-side discovery path and source-of-truth cross-link.

[LauraGPT]

FunClip integration is now merged and released:

• FunClip PR: feat: add MOSS diarization backend modelscope/FunClip#207
• FunClip v2.2.0: https://github.com/modelscope/FunClip/releases/tag/v2.2.0
• Production guide: https://www.funasr.com/en/deploy/moss-transcribe-diarize.html

The shipped path uses the tested vLLM /v1/audio/transcriptions JSON contract from FunASR 1.4.9, preserves MOSS S01/S02 segments for SRT and speaker clipping, keeps short speaker turns, and rejects truncated final segments. It deliberately does not attach external VAD or speaker models. OpenMOSS ownership and the segment-level timestamp boundary are called out in both READMEs and release notes.

Live validation used the pinned revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8 on an H100 with a two-speaker WAV.

[LauraGPT]

FunClip v2.2.0 is now also documented on the deployed FunASR product site:

• English: https://www.funasr.com/en/blog/funclip-v2-2-0-moss-speaker-clipping.html
• Chinese: https://www.funasr.com/blog/funclip-v2-2-0-moss-speaker-clipping.html
• Release: https://github.com/modelscope/FunClip/releases/tag/v2.2.0

The guide explicitly identifies MOSS-Transcribe-Diarize as an OpenMOSS third-party model, pins revision e8681d68e7042738ffca8ac8212bc8fcb1131ab8, and limits the supported FunClip path to the live-tested vLLM integration with segment timestamps, speaker SRT, and speaker clip export.

[yhzx233]

We really appreciate the effort you put into bringing MOSS-Transcribe-Diarize into the broader ecosystem!