Clipsai

Clips AI's appeal is structural rather than cosmetic. Most clip tools sell you a hosted editor and a render queue; Clips AI hands you two functions and gets out of the way. Three lines of setup produce a Transcription object, ClipFinder turns that into clips with start_time and end_time, and resize() returns crop segments that track the active speaker. That is the entire product surface, and for an engineering team it is the right amount. The strengths are concrete. Transcription runs through WhisperX, which adds word-level start and stop times on top of Whisper, and those word timings are what let ClipFinder place boundaries from the transcript rather than from arbitrary intervals. Reframing uses Pyannote for speaker diarization inside resize(), which is why a wide two-person interview can come back as a vertical clip that follows whoever is talking. The aspect_ratio argument is a tuple, so (9, 16) is the documented example rather than a hard limit. And because it is a library, the clipping algorithm is yours to fork — swap in your own boundary logic without waiting on a vendor roadmap. The costs are equally concrete. Setup is multi-step: pip install clipsai, then WhisperX from its GitHub repository rather than PyPI, then libmagic and ffmpeg on the system. Resizing needs a Hugging Face access token because of Pyannote, one more step before resize() will run (the docs note you are not charged for it). Clips come back as data, so trimming, captioning, reviewing, and publishing are all downstream work you own. Where it fits: engineering teams with a repurposing pipeline, podcast and interview shops producing many vertical clips from long episodes, and anyone who needs footage to stay on their own infrastructure. Where it doesn't: anyone who wants a GUI, a review step, or clips from footage whose signal isn't speech.