FunClip 2.2.0: speaker-aware video clipping with MOSS / 基于 MOSS 的说话人视频剪辑

[LauraGPT]

FunClip 2.2.0 adds an opt-in MOSS-Transcribe-Diarize path for long videos with multiple speakers. The third-party OpenMOSS model produces transcription, segment timestamps, and anonymous speaker labels (S01, S02, ...); FunClip turns those segments into speaker-aware SRT and per-speaker audio/video clips.

FunClip 2.2.0 新增可选的 MOSS-Transcribe-Diarize 路径，面向长视频和多人对话。第三方 OpenMOSS 模型输出转写、分段时间戳与匿名说话人标签（S01、S02……），FunClip 可据此生成带说话人标签的 SRT，并按说话人导出音频或视频片段。

Start / 启动

Run a local or remote vLLM transcription service with the pinned MOSS model, then start FunClip:

git clone --branch v2.2.0 https://github.com/modelscope/FunClip.git
cd FunClip
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. Use --moss-base-url for another endpoint; supply credentials only through MOSS_API_KEY.

默认服务地址为 http://127.0.0.1:8898/v1，其他地址可通过 --moss-base-url 指定；凭据仅通过 MOSS_API_KEY 环境变量传入。

What is supported / 支持范围

• speaker-aware SRT using model-provided S01, S02, ... labels;
• clipping all turns from one selected speaker, including turns shorter than one second;
• explicit failure when the last MOSS segment is truncated, instead of silently dropping output;
• pinned model/runtime guidance and SHA256-verified FunClip source archives.

MOSS timestamps are segment-level. Speaker clipping and SRT are supported, while precise arbitrary-text clipping remains on Paraformer. Do not attach an external VAD or speaker model: pre-chunking can break speaker-label consistency.

MOSS 提供的是分段级时间戳，适合说话人剪辑和 SRT；需要按任意精确文本剪辑时仍使用 Paraformer。不要再外挂 VAD 或说话人模型，因为预切分可能破坏全局说话人标签的一致性。

Evidence and ownership / 验证与归属

The release content is exact commit c205bf32a8b11226ff5e8acb9a3c7a1f00cd3b06. Its suite completed with 84 passed and 1 skipped. A pinned H100/vLLM probe produced two speaker segments, valid SRT, and the expected S02 clip.

MOSS-Transcribe-Diarize remains an OpenMOSS model under its upstream license. FunClip integrates the public serving contract and does not bundle the model weights. The labels are anonymous diarization labels, not enrolled-person identity verification.

MOSS-Transcribe-Diarize 仍是 OpenMOSS 所有并维护的第三方模型；FunClip 仅接入其公开服务契约，不打包模型权重。这些标签是匿名说话人分段结果，不是已注册人员身份验证。

Links / 入口

• FunClip v2.2.0 release
• 中文完整指南
• English guide
• FunASR MOSS deployment guide
• OpenMOSS upstream

Please report the exact FunClip tag, MOSS revision, vLLM version, GPU, source-video duration/languages, and a redacted segment excerpt. Reports will remain open for real-environment confirmation instead of being closed merely because the release exists.

反馈时请附 FunClip tag、MOSS revision、vLLM 版本、GPU、源视频时长/语言和脱敏后的分段输出。相关问题会等待真实环境复测，不会因为版本已经发布就直接关闭。