FunClip v2.2.0

MOSS speaker-aware clipping

FunClip v2.2.0 adds an opt-in --model moss path for the third-party OpenMOSS MOSS-Transcribe-Diarize model through a local or remote vLLM transcription service. OpenMOSS owns and maintains the model; this release integrates its public serving contract and does not bundle model weights.

What is included

• long-form ASR, speaker identity, and segment timestamps without an external VAD or speaker model
• SRT generation with spkS01, spkS02, and other model-provided speaker IDs
• speaker-based audio/video clipping, including turns shorter than one second
• explicit failure on a truncated final MOSS segment instead of silently dropping partial output
• environment-only bearer credential handling through MOSS_API_KEY
• pinned model revision and bilingual deployment guidance

Start

pip install -U -r requirements.txt
python funclip/launch.py --model moss --moss-backend vllm

The default service URL is http://127.0.0.1:8898/v1; override it with --moss-base-url. See the production guide.

Boundaries

MOSS timestamps are segment-level. SRT, speaker clipping, and LLM timestamp clipping are supported; precise arbitrary text clipping remains on Paraformer. Do not attach an external vad_model or spk_model, because pre-chunking breaks global speaker identity.

Validation

The exact release content is commit c205bf32a8b11226ff5e8acb9a3c7a1f00cd3b06. The suite completed with 84 passed and 1 skipped. A live H100/vLLM test produced two speaker segments, valid SRT, and the expected S02 clip. Verify downloaded assets with SHA256SUMS.

Source PR: #207
OpenMOSS integration context: OpenMOSS/MOSS-Transcribe-Diarize#48

Deployment guide

• Chinese · English
• The guide keeps OpenMOSS ownership explicit and documents the validated vLLM boundary, segment-level timestamps, speaker SRT/clips, pinned model revision, and archive checksums.