Key Takeaways
- FunClip (github.com/modelscope/FunClip) is a free, open-source, self-hosted video clipping tool β not a hosted web service
- Developed by ModelScope, Alibaba's open-model platform, as part of its TONGYI speech lab work
- MIT licensed for the source code, confirmed via the GitHub repository's LICENSE file; underlying FunASR model weights are separately licensed under Apache 2.0
- Speech-to-text via FunASR's Paraformer-Large model, with CAM++ speaker recognition and SeACo-Paraformer hotword customization
- LLM-assisted clip selection: describe the segment you want and an LLM backend identifies matching timestamps
- Interfaces: a local Gradio web UI (default at
localhost:7860) and a command-line script,videoclipper.py - Language support: Mandarin Chinese by default, an English mode via a
-l enflag, and multilingual support (plus emotion detection) through the SenseVoice model - Latest release at the time of this review: v2.2.1
- Over 6,300 GitHub stars as of this review
π In One Sentence
FunClip is a free, open-source (MIT), self-hosted video clipping tool from ModelScope that transcribes video with FunASR speech-recognition models and uses an LLM to find and cut the segments you want, with automatic subtitle generation, speaker recognition, and over 6,300 GitHub stars.
π¬ In Plain Terms
FunClip watches a long video for you, writes down everything that is said (transcription), and then, when you describe what moment you want, an AI model finds that moment and cuts the clip automatically, with subtitles already burned in or attached. It runs on your own computer or server through a local web page or the command line β it is free, but you need to install it yourself, and the AI parts need enough hardware to run speech and language models.
πNote: This review is based on FunClip's own GitHub repository and README. It does not claim PromptQuorum has run hands-on transcription-accuracy benchmarks of FunClip, and it presents ModelScope/Alibaba's role as the developer factually, without geopolitical framing.
What Is FunClip?
FunClip is a self-hosted, open-source tool that automates video clipping by combining speech-to-text transcription with LLM-assisted moment selection, rather than requiring you to scrub through raw footage manually. It is developed by ModelScope, Alibaba's open-model platform, and builds on that team's own FunASR speech-recognition models.
- Product type: a self-hosted, web/CLI tool β installs via git clone and Python, not a downloadable desktop app or hosted SaaS product
- Developer: ModelScope (Alibaba's open-model platform), specifically its TONGYI speech lab work
- License: MIT for the FunClip source code itself, confirmed via the GitHub repository's LICENSE file; the underlying FunASR model weights it depends on are separately licensed under Apache 2.0
- Core dependency: FunASR, ModelScope's own open-source speech-recognition toolkit, which FunClip uses for transcription rather than building its own ASR model from scratch
- Scale: over 6,300 GitHub stars as of this review
- Latest release: v2.2.1, which the project's own changelog describes as preserving selected subtitle colors through a Pillow-based renderer
What Does FunClip Actually Do?
FunClip takes a video file, transcribes the speech in it, and then lets you (or an LLM, based on your description) select specific moments to cut into standalone clips with subtitles already generated.
- Automatic transcription: FunASR's Paraformer-Large model transcribes the video's speech with integrated timestamp prediction, so each spoken segment is mapped to an exact point in the video
- Speaker recognition: CAM++ speaker identification lets you clip segments by a specific speaker, useful for interviews, panels, or multi-person recordings
- Hotword customization: SeACo-Paraformer integration lets you specify entity names or specific terms in advance to improve recognition accuracy for those words
- LLM-assisted clip selection: instead of manually scrubbing the full transcript, you can describe what moment you want and an LLM backend identifies the matching timestamps to cut
- Automatic subtitle generation: FunClip generates SRT subtitles for both the full video and each individual clipped segment
- Subtitle rendering: version 2.2.1 added a Pillow-based renderer that preserves selected subtitle colors in the output video
- Multi-segment output: FunClip can generate multiple clips from a single source video in one pass, rather than requiring a separate run per clip
Usage Examples: Two Ways to Use FunClip
These are workflows built from FunClip's own documented features β not hypothetical use cases.
Platform, Pricing, and Licensing
Platform
- What FunClip states:
- Self-hosted Python application; a local Gradio web UI plus a command-line script. Cross-platform in principle, though specific OS support is not enumerated in the docs.
Cost
- What FunClip states:
- Free and open source. No paid tier or hosted service; you run it on your own hardware.
Licensing
- What FunClip states:
- MIT for the source code, confirmed via the GitHub repository's LICENSE file; FunASR model weights are separately licensed under Apache 2.0.
Install method
- What FunClip states:
- Clone the GitHub repository and install dependencies from
requirements.txtinside a Python 3.12+ virtual environment, per the project's own README. Versioned release archives (for example, FunClip-2.2.1.tar.gz/.zip) are also available with SHA256 checksums.
Verify the current recommended Python version and dependency list directly in the GitHub repository before installing, since requirements can change between releases.
Install FunClip
FunClip installs free via git clone and pip, and its source code and versioned release archives are on GitHub.
Source | Link |
|---|---|
| GitHub repository (source code, MIT) | github.com/modelscope/FunClip |
| Versioned release archives (with SHA256 checksums) | github.com/modelscope/FunClip/releases |
| FunASR (underlying speech-recognition toolkit) | github.com/modelscope/FunASR |
| ModelScope platform | modelscope.cn |
FunClip requires a Python 3.12+ environment and enough local hardware to run FunASR's speech-recognition models plus your chosen LLM backend β there is no GUI installer and no hosted free tier to sign into.
FunClip vs. Whisper.cpp
FunClip and whisper.cpp both run speech-to-text locally, but they solve different problems. Whisper.cpp is a lean, dependency-light C/C++ transcription engine you embed in other projects; FunClip is a full application layer on top of FunASR transcription, adding LLM-assisted clip selection, speaker recognition, and subtitle-burned video export.
Aspect | FunClip | Whisper.cpp |
|---|---|---|
| Primary purpose | End-to-end video clipping app built on transcription | A transcription engine, embeddable in other tools |
| Speech model | FunASR Paraformer-Large / SenseVoice | OpenAI Whisper weights (C/C++ port) |
| Output | Cut video clips with generated subtitles | Transcript/subtitle text only |
| Interface | Local Gradio web UI + CLI script | Command-line and library bindings |
| Runtime dependency | Python 3.12+, FunASR, an LLM backend | No Python runtime required |
| Developer | ModelScope (Alibaba) | Georgi Gerganov / ggml-org |
If your goal is finished video clips with subtitles and minimal manual editing, FunClip's built-in clip-selection and export pipeline does more out of the box. If you only need raw transcription text or subtitles to feed into your own pipeline, without Python or a full video-editing layer, see the whisper.cpp review. Verify current feature sets on each project's own repository before choosing.
Who Should Use FunClip?
FunClip fits people who need to turn long recordings into short clips regularly and want that process automated by transcription plus an LLM, rather than manual scrubbing.
What FunClip Is Not Good For
FunClip is not a good fit if you want a no-install hosted service or a general-purpose video editor beyond clip selection and subtitles.
- Not a hosted product β there is no managed cloud tier; you install and run it yourself via git clone and a Python environment
- Not a general video editor β it is focused specifically on transcription-driven clip selection and subtitle generation, not timeline editing, color grading, or effects
- Not zero-configuration β getting LLM-assisted clip selection working requires configuring an LLM backend yourself, on top of the Python/FunASR install
- Not guaranteed to have enumerated OS-level platform support β the project's own documentation does not list specific supported operating systems beyond describing it as a Python/Gradio app, so verify compatibility with your own environment before relying on it
- Not independently benchmarked by PromptQuorum for transcription accuracy β this review is based on FunClip's own GitHub repository and README, not hands-on accuracy testing
Common Mistakes When Evaluating FunClip
Most confusion about FunClip comes from expecting a full video editor, missing the language-mode flag, or assuming it needs no additional LLM setup.
Competitors and Alternatives
FunClip is most often compared to other local, self-hosted speech-to-text tools such as whisper.cpp, faster-whisper, and MacWhisper β its main differentiator is going beyond raw transcription to full LLM-assisted clip selection and subtitle-burned video export.
Tool | Best known for | Link |
|---|---|---|
| whisper.cpp | Free, MIT-licensed C/C++ port of OpenAI Whisper for on-device transcription | whisper.cpp review |
| faster-whisper | CTranslate2-based reimplementation of Whisper optimized for speed on GPU/CPU | faster-whisper review |
| MacWhisper | Polished native macOS transcription app built on Whisper models | MacWhisper review |
This list reflects tools commonly compared to FunClip on the transcription side, not an independent PromptQuorum ranking β verify each tool's current feature set before choosing. None of these three add FunClip's LLM-assisted clip-selection and video-export layer; they are transcription engines/apps, not clipping tools. See also the dedicated FunClip vs. Whisper.cpp comparison above.
Frequently Asked Questions
What is FunClip?
FunClip (github.com/modelscope/FunClip) is a free, open-source, self-hosted video clipping tool developed by ModelScope that transcribes video with FunASR speech-recognition models and uses an LLM to find and cut the segments you want.
Is FunClip free?
Yes, FunClip is free and open source (MIT-licensed source code; FunASR model weights are separately licensed under Apache 2.0). There is no paid tier or hosted service β you run it on your own hardware.
How do I install FunClip?
Per the project's own README, clone the GitHub repository and install dependencies from requirements.txt inside a Python 3.12+ virtual environment. Versioned release archives with SHA256 checksums are also available on the releases page.
Does FunClip support English?
Yes, FunClip defaults to Mandarin Chinese but documents an English mode via a -l en flag, plus multilingual support through the SenseVoice model.
Does FunClip work from the command line?
Yes, in addition to the local Gradio web UI (default at localhost:7860), FunClip provides a command-line script, videoclipper.py, for direct video recognition and clipping.
Does FunClip generate subtitles automatically?
Yes, FunClip automatically generates SRT subtitles for both the full video and each individual clipped segment, using FunASR's transcription with integrated timestamps.
Who develops FunClip?
ModelScope, Alibaba's open-model platform, develops FunClip as part of its TONGYI speech lab work. The canonical GitHub repository is modelscope/FunClip.
Does FunClip need an LLM to work?
The core transcription and manual clip-export features work without an LLM. The LLM-assisted clip-selection feature, where you describe a moment and the tool finds it, requires configuring an LLM backend separately.
Can FunClip identify different speakers in a recording?
Yes, FunClip includes CAM++ speaker recognition, letting you clip segments by a specific speaker β useful for interviews, panels, or multi-person recordings.
Has PromptQuorum independently tested FunClip's transcription accuracy?
This review is based on FunClip's own GitHub repository and README, rather than hands-on transcription-accuracy benchmarking by PromptQuorum.