FunClip
modelscope/FunClip · Python · MIT
Transcribe, subtitle and LLM-clip long videos — locally.
FunClip runs FunASR/SenseVoice locally to transcribe video with word-level timestamps, then lets you (or an LLM) pick segments and export properly cropped clips with hardcoded subtitles. No paid API needed for the ASR path.
Score
Setup at a glance
git clone https://github.com/modelscope/FunClip.git && pip install -r requirements.txt Deep dive
FunClip is Alibaba ModelScope's video-clipping workbench: it transcribes locally with the industrial-grade Paraformer-Large ASR model, then lets an LLM pick the clip-worthy moments — no paid transcription API involved.
Recognition quality is where the engineering went: SeACo-Paraformer hotword customization sharpens domain vocabulary, CAM++ speaker recognition tells speakers apart, and the whole interaction runs on a Gradio UI that starts with one command.
The same pipeline is available as a ModelScope Studio and a Hugging Face Space if you would rather try it in the browser before self-hosting.