How Clippen turns one long recording into vertical clips

What happens to your video after you submit it?

You hand Clippen one long recording: paste a link or upload the file. From that point the work happens on our servers, not in your browser. You can close the tab, and the clips are waiting when you come back.

The job runs as a queue of six stages, in a fixed order, because each one needs the output of the one before it. The transcript has to exist before a model can judge which sixty seconds are worth keeping. The scene boundaries have to exist before a cut can be placed somewhere that is not the middle of a shot. The selected moments have to exist before there is anything to reframe.

Nothing is published anywhere as part of processing. Clips are yours to download, and posting to TikTok, Instagram Reels or YouTube Shorts happens only when you connect an account and press post. Clippen is built for clipping video you own or have the right to use — your own podcast, your own stream, your own talk. You keep every right you already had in the source and in the clips; Clippen takes no licence beyond processing the file you submit. Clipping someone else's upload is between you and the destination platform: YouTube's inauthentic content policy of July 2025 names recycled clips explicitly.

The six stages, in order

Transcription with word-level timestamps

faster-whisper transcribes the source audio and returns a timestamp for every individual word, not just every sentence. Word-level timing is what lets a clip start on a word boundary instead of half a syllable in, and it is what makes karaoke captions land on the right frame. The spoken language is detected automatically.

Scene boundary detection

PySceneDetect scans the video for cuts and hard visual transitions. Those boundaries stop a clip from opening mid-shot, which is the most common reason an auto-generated clip reads as machine-made.

Moment selection

Google Gemini receives the timestamped transcript and the scene boundaries together and returns the strongest self-contained segments of 15 to 60 seconds each, judged on how hard the opening lands, whether the segment pays off, and whether it still makes sense to somebody who has not heard the rest of the episode. The count scales with the material rather than being fixed: an hour of conversation typically yields 6 to 12.

Vertical reframing

Each clip is recomposed from 16:9 to 9:16. TRACK follows a single speaker with MediaPipe face detection, damped by a stabiliser that holds the camera still inside a safe zone instead of chasing every head movement. GENERAL keeps the full width over a blurred backdrop for group shots and landscapes. SPLIT stacks two speakers in half-frames when both are genuinely in the same shot. SCREENCAST and INSET handle slides, shared screens and stream layouts where the meaning lives outside the centre of the frame.

Captions, hooks and pacing

Captions are rendered word by word from the transcript and burned in, with the spoken word highlighted as it lands. An AI-written hook headline covers the opening seconds, where a short-form viewer decides whether to stay. Optional punch-ins push in on the beats of the clip so a static camera does not read as a static clip.

Export and publishing

Clips come out at 1080×1920 in 9:16, 1:1 and 16:9. Download them, or post to TikTok, Instagram Reels and YouTube Shorts from the dashboard. ElevenLabs dubbing can translate a clip into 31 languages while keeping the speaker's voice, and the dubbed audio is re-transcribed so the captions match the new language.

How long does it take?

Processing runs at 0.30 to 0.36× real time, measured end to end on real sources rather than on a short demo file. In practice that means a 60-minute podcast episode is finished in about 20 minutes, a 20-minute conference talk in about seven, and a 90-minute webinar in a little under half an hour.

The number is worth stating precisely because it changes how you plan. Clipping an hour-long episode by hand — finding the moments, cutting them, reframing each one to vertical, typing and timing the captions — is most of a working day, and it is the reason most people's back catalogue is sitting unpublished. Twenty minutes of unattended processing is a different kind of decision. You start the job at the end of the recording session and the clips exist before you have finished tidying up.

Almost all of that time is transcription and video encoding. The model call that picks the moments is a small fraction of it, which is why a longer source costs more time roughly in proportion to its length rather than in sudden jumps.

Why do clips not start in the middle of a sentence?

Because two separate signals are used to place the cut, and each one catches what the other misses.

The transcript carries a timestamp on every individual word, not on every sentence. That is what makes it possible to start a clip exactly on the first consonant of the first word rather than a fraction of a second into it, which is the audible tell that a clip was cut by a machine. Word-level timing is also what lets the captions highlight each word as it is actually spoken.

Scene detection supplies the other half. A transcript alone can find a brilliant sentence and have no idea that the camera cuts halfway through it. Scene boundaries alone can find clean visual cuts with nothing worth saying between them. The model that selects moments receives both together, so it can choose a segment that is quotable and visually intact at the same time.

The selection is also judged on whether the segment stands on its own. A sixty-second passage that opens with "and that is exactly why I disagree" is useless to a viewer who has not heard the previous ten minutes, however good the line is. Self-containment is a selection criterion, not an afterthought.

How does the vertical crop decide what to keep?

A 16:9 frame recomposed to 9:16 loses about three quarters of its width, so something has to decide what that width was for. Clippen picks a layout per scene rather than applying one crop to the whole video.

• TRACK follows a single speaker using MediaPipe face detection. The camera is damped: it holds still inside a safe zone and only moves when the subject leaves it, because a crop that chases every head movement swings visibly and reads as amateur even when the framing is technically correct.
• GENERAL keeps the full width over a blurred backdrop, for group shots and landscapes where cropping to any one point throws away the subject.
• SPLIT stacks two speakers in half-frames, but only when both faces are genuinely in the same shot for a sustained part of the scene. That test is what separates a real two-shot from a shot/reverse-shot edit, where stacking would show the same person twice.
• SCREENCAST and WIDE handle slides and shared screens. Content that spans most of the frame cannot survive a crop, so it is stacked over the presenter or kept full width instead of being cut away.
• INSET handles the stream layout: a full-width screen with a small webcam box in a corner. The screen goes on top and the webcam is enlarged underneath.
• Speaker cuts and punch-ins sit on top of the layout: hard cuts to whoever is talking, and a gentle push in on the beats, so a static camera does not produce a static-looking clip.

What does a run cost?

Clippen is a paid tool and bills by the length of what you feed it: one credit is one minute of source video, and that one charge covers the whole run — every clip the video yields, the vertical reframing, the captions and the hook. A 60-minute podcast costs 60 credits whether it comes back with six clips or twelve. Plans start at $9.99 a month for 50 credits ($4.99 a month billed annually) and run to $199.99 a month for 1200 credits. There is no free tier.

The unit matters when you compare tools. Billing per minute of source means the cost of an episode is knowable before you press start: a 45-minute interview is 45 credits, whatever comes back. Billing per finished clip means the cost moves with an output you cannot predict.