Let the transcript hold the scissors

What it does FunClip is a locally deployable video clipping tool that transcribes footage using Alibaba’s open-source FunASR models, including Paraformer-Large and SeACo-Paraformer. It surfaces the transcript with timestamps in a Gradio interface; you highlight text or pick a speaker, and it slices the matching video segments. It also generates SRT subtitles for both the full video and each clip.

The interesting bit The project treats the transcript as the editing timeline. Its newer LLM integration feeds the generated SRT subtitles to models like Qwen or GPT with a prompt, then parses the model’s output to extract timestamps and automate the cut list. It is essentially well-organized glue between ASR, speaker diarization, and an LLM API—but for a task this tedious, the glue is the product.

Key highlights

• Built on Alibaba’s open-source Paraformer-Large and SeACo-Paraformer ASR models with timestamp prediction and hotword customization
• Speaker-aware clipping via the CAM++ model, letting you isolate specific speakers by ID
• LLM-assisted “AI clipping” that uses language models to suggest segments based on subtitle content
• Expanded language support through Fun-ASR-Nano (31 languages) and SenseVoice for emotion/audio event detection
• Core pipeline runs fully offline; only the optional LLM features call out to external APIs

Caveats

• The LLM clipping mode requires bringing your own API key; there is no bundled local language model.
• Burning subtitles into clips requires a separate ImageMagick installation.
• A few quality-of-life features—like reversing period selection and silence removal—remain on the roadmap.

Verdict Worth a look if you routinely pull quotes or isolate speakers from long footage. Skip it if you need frame-perfect precision or already have a polished NLE workflow you love.