Two people sitting across from each other end up at opposite edges of a 16:9 frame, and a vertical crop keeps neither of them. ClipSpeedAI stacks the speakers instead of choosing between them — a two-pane split, a 3-up or a 4-up grid — with tracking inside each pane so faces stay centred while people move.
Real ClipSpeedAI output — every clip below was scored, cut, reframed to 9:16 and captioned automatically from the video above it.
Source video
15:27
You do not draw masks, set keyframes, or nominate who goes on top.
Paste the YouTube URL of the episode or upload the recording — a Zoom export, a Riverside session, a conference panel, a co-stream VOD. What matters is that it is the finished combined video with everyone visible in the frame. Two hours per file on paid plans, thirty minutes on the free demo.
The engine detects how many people are actually in the picture across the moment it is cutting. Two faces get a split, three get a 3-up, four get a grid, and a stretch where only one person is in shot gets a single full-frame crop instead of an empty pane. The footage decides, not a setting you picked at the start.
Each cut arrives at 9:16 with the panes stacked, word-by-word captions burned in, and a 0-100 score. If a particular moment would play better as a single speaker filling the frame, switch the layout on that clip and render it again.
Seven layouts ship in total. These are the ones that exist because conversations do not fit in a vertical rectangle.
The workhorse for interviews and two-host shows. Both faces sit in their own pane, one above the other, each cropped tight enough to read on a phone. The reason this beats cutting between speakers is that a listener’s reaction is often the better half of the moment, and a single-speaker cut throws it away.
Three panes in a vertical column, which is roughly the practical limit for faces that stay readable at phone size. It suits panels where one person is steering and two are responding, since the eye can follow a three-way exchange without the frame having to cut at all.
A two-by-two arrangement for four-person recordings and remote roundtables. Each tile is smaller, so this works best on close-framed webcam footage where heads already fill their source frames and badly on wide room shots where everyone is a distant figure at a table.
A pane is not a static crop of a fixed rectangle. Each one follows its speaker, so somebody leaning back, turning to a second camera or gesturing off to one side stays framed. Movement is smoothed rather than snapped frame to frame, because a crop that twitches on every small motion is more distracting than a slightly loose one.
Forcing a split onto footage that only ever shows one person produces a black rectangle where the second speaker should be, which is the most common way automated vertical conversion fails visibly. Layout is selected from what is actually in the frame during that clip, and it is changeable per clip afterwards.
Panels that pull up a slide, a chart or a demo get the screenshare layout instead: the shared display takes the readable portion of the frame with the speaker kept visible below it. Losing the slide is as bad as losing the speaker when the whole point of the clip was the number on screen.
Gameplay with two or more people on cam is a different geometry problem, and the PiP layout keeps the action large with the reactions inset. This is the layout you want for a clip where what happened in the game and how people responded are equally the story.
Text and stacked panes are competing for the same vertical inches. Word-by-word captions are burned in across eleven styles, and on a 3-up or 4-up the plainer styles read better than the heavy ones simply because there is less room. Captions carry most of the comprehension on a muted feed, so they get priority over decoration.
Multi-speaker clips get the same treatment as everything else: cut on sentence boundaries so a clip never opens mid-word, filler and dead air removed — which matters more in conversation because turn-taking generates gaps — and a 0-100 grade to rank the batch.
Connect a YouTube, Twitch or Kick channel and multi-speaker clips are produced while the broadcast is still running. A three-way argument on a live show is exactly the kind of moment that is worth far more posted within the hour than the next morning.
The guest says something and the host reacts, and both halves are the clip. See the podcast clip generator for the wider workflow around episode-length recordings.
Four chairs on a stage in one wide shot is nearly unusable vertically. Cropping each speaker into their own pane is what makes a panel postable at all.
Disagreement is the format, and disagreement is unreadable if you can only see whoever is talking. Both faces at once is the whole point.
Grid-view calls already put everyone in a box, and restacking those boxes vertically converts a call recording into something that works on a phone.
Two webcams over gameplay, or two friends reacting to the same thing. Both cams stay in shot instead of the clip picking a side.
Multi-host webinars and internal panels produce long recordings with a handful of quotable exchanges buried in them, and those exchanges usually involve at least two people.
What a conversation loses to a vertical crop is the other person. Seat two people across from each other and they occupy opposite edges of a wide frame, so the strip that survives is the gap in between — a bookshelf, a mic arm, a doorway, a lamp. Neither face is in it. It is the single most recognisable failure of naive auto-cropping, and once you have seen it you cannot unsee it on other people’s clips. If you want the pixel-by-pixel version of how little of a wide frame comes through, it is worked out in full on the vertical video converter page; the part that matters here is simpler, which is that a two-hander gives you no version of the shot where cropping costs you nobody.
The next attempt is usually to track the active speaker and cut between them. That is better, and for a monologue it is correct. In a conversation it quietly removes the thing that made the moment work. Someone delivers a line and the other person’s face does something, and if the clip is on the speaker for all forty seconds the reaction never happened as far as the viewer is concerned.
Stacking is the answer because vertical has height to spare. A 9:16 frame is tall enough for two cropped faces at a size that reads fine on a phone, and the format is watched close to the eyes on a small screen where smaller faces still work. What is scarce in vertical is width, and a split screen spends the dimension you have on the problem you have.
The first requirement is that each pane keeps its person. If tiles reshuffle mid-clip, a viewer has to re-learn the arrangement every few seconds and the clip becomes exhausting to watch, no matter how good the exchange is. A stable arrangement is more important than an optimal one.
The second is tracking that does not fidget. Cropping tight to a face means small head movements translate into large frame movements, and a pane that re-centres on every micro-motion reads as a camera operated by somebody nervous. Following the subject with smoothing applied is the difference between a crop you notice and one you do not.
The third is knowing when a pane should not exist. Not every second of a four-person recording has four people visible — a director cuts to a close-up, someone leaves to get coffee, the camera pushes in on the speaker. A layout that insists on four tiles will render empty boxes. Choosing the layout from what is actually in the frame during that specific clip is the fix.
The fourth is leaving room for text. Captions are what most of the audience is reading, since the majority of feed viewing happens with the sound off. A layout that fills every pixel with faces and then drops caption text on top of somebody’s mouth has solved the framing problem by creating a legibility one.
Most of the quality of a multi-speaker clip is decided before you press record. If you already know panels are going to be clipped, a few habits at capture time are worth more than anything a tool can do afterwards.
Frame each person tighter than feels right for the landscape edit. A wide two-shot with plenty of headroom looks composed on a monitor and produces small, distant faces once it is cropped into a pane. Faces that already fill their portion of the source frame survive the crop with resolution to spare.
Keep everyone in shot rather than cutting to close-ups in the master edit. Every second where the director has punched in on one speaker is a second where no split screen is possible, because the other person genuinely is not in the pixels. If you publish both a directed cut and a raw multi-view recording, submit the raw one.
Light both sides of a remote conversation. It is common for a host in a studio and a guest on a laptop webcam to end up in the same grid, and the difference is jarring at pane size. Asking a guest to sit facing a window costs nothing and closes most of the gap.
Finally, leave a beat after a good line before jumping in. Overlapping speech is the hardest thing to clip cleanly, and the reaction you want in the second pane needs a moment of quiet to be visible in. Interviewers who learn this get noticeably better clips out of identical conversations.
It composites from your finished video, not from separate camera files. If you recorded isolated per-speaker tracks in Riverside, Zoom or an equivalent, export the combined edit and submit that. There is no multitrack ingest that takes four ISO files and arranges them for you, and if that is what you need then a full NLE is the right tool.
Four panes is the ceiling. A six-person panel does not become a 6-up, because six faces on a phone screen are too small to carry any expression. What you get instead is the people who are actually speaking during that clip, which is usually the honest edit anyway.
It cannot show a person who is not in the picture. If your source is single-camera and the director cut to a close-up of the host for that whole exchange, the guest does not exist in those pixels and no layout can retrieve them. Multi-speaker layouts work on multi-speaker footage, which sounds obvious but is the most common mismatch we see.
And wide room shots are a genuinely hard case. If four people are sitting at a distant conference table, cropping each of them into a small tile gives you four low-resolution figures rather than four faces. Close-framed webcams and proper podcast setups produce far better results than a single camera at the back of a room.
It is worth being precise about what you are giving up, because the honest answer is not nothing. A person cutting a panel by hand breaks the split deliberately: they hold both faces through the setup, then punch in to full frame on the exact syllable where the reaction lands, then pull back out. That rhythm — two panes, one pane, two panes — is the thing a good conversation editor is actually doing, and the layout here is chosen once per clip rather than changed inside it.
A human also sizes panes unevenly. Sixty per cent of the height to whoever is carrying the exchange and forty to the listener reads more naturally than an even split, and an editor adjusts that per moment. Panes here are equal by construction. That is a real trade and not a subtle one on a two-hander where one person talks for most of the clip.
What swings the other way is that the engine does the same thing on clip fifteen as it did on clip one. Say you run a weekly three-person show and pull eight clips from every episode: that is four hundred clips a year, each needing masks, positions and keyframes whenever somebody shifts in their chair. Nobody sustains hand-built splits at that rate, and the version that actually happens is not the careful edit — it is three careful clips and five that never got made.
The biggest mistake here is deciding this in the abstract. Cut the batch, find the one clip that genuinely deserves an hour, and rebuild only that one in a timeline. Everything you learn doing it makes the next batch easier to review, and the other seven clips went out on Tuesday instead of never.
Panes are a solution to a specific problem — several people, one wide frame, a vertical output — and there are several moments where the alternatives beat them outright.
A single tracked crop of whoever is speaking is the strongest option more often than people expect. Any moment that is really a monologue with an audience, any stretch where one answer runs a full thirty seconds, any clip where the other person is politely nodding: a full-frame face reads better than a half-frame face plus a half-frame of nodding. It turns out this is also the right call on wide room shots, because one legible face beats four unreadable tiles.
Keeping the original 16:9 is an alternative nobody considers because vertical has become the default. For YouTube, LinkedIn and an embed on your own site the landscape frame already holds everyone at full size and needs no layout decision at all. A 90-minute conference panel destined for a company blog does not need to be vertical, and forcing it there costs resolution for no gain.
For gameplay and co-streams, picture-in-picture usually beats an even split, since the action and the reaction are not equally important and PiP says so in the composition. For a moment built around a slide or a chart, the screenshare layout is the alternative: the thing being shown is the content and the faces are context.
And if you have isolated per-speaker camera files, a full editing suite with a multicam workflow will beat any automatic layout, because it is cutting between full-resolution sources instead of cropping regions out of one composite. That is a different amount of work, but if the deliverable is one flagship clip a month rather than eight a week, it is the better tool and we would rather say so.
Most disappointing multi-speaker clips are not broken, they are mismatched — the layout is doing something reasonable for footage other than yours. Four symptoms account for nearly all of it, and three have a fix that takes one re-render.
The split picked the wrong pair. On a five or six-person panel the two panes go to the people carrying that stretch of talk, so somebody who lands one sharp interjection can lose their pane to whoever simply spoke more seconds. If the interjection is the reason the clip exists, override that clip to a 3-up and render it again — three panes will hold all the participants who matter, and the person you wanted is back in the frame.
The clip opens on the listener. Conversations overlap at the edges, and a cut placed on a clean sentence boundary can still begin with the tail of the previous turn while the camera is already on the person about to answer. Nudge the in point a beat later. It is a two-second adjustment and it is the single most common reason a good exchange feels like it starts in the middle of something.
The framing is fine but the clip drags. Filler and dead air have already been taken out by this point, so what is left is real conversational time — thinking, nodding, a pause before an answer. A split doubles how much of the screen is visibly doing nothing during those beats. Either tighten the in and out points around the exchange itself, or drop to a single crop, which hides the waiting instead of framing it twice.
One pane looks worse than the other. That is a studio host next to a guest on a laptop webcam, and no layout corrects exposure or lens quality. What does help is spending fewer panes: a two-pane split gives the weaker source considerably more width than a 4-up tile does, so if the choice is available on that clip, take the split.
Observations from operating the pipeline in production — not general advice.
Face count is not a constant inside a thirty-second exchange. A directed edit will punch in on the host, hold there four seconds, then cut back out. If the layout decision is taken from a single frame, it can land inside that punch-in and render the whole clip as a solo crop with the guest absent from their own answer. Reading detections across the entire span being cut is what avoids that, and the reason it matters far more on panels than on a lone presenter is that a multi-camera conversation changes its composition several times inside a window a talking head would fill with one unbroken shot.
We once shipped a long stretch of clips where face detection had quietly stopped reporting and every crop fell back to the centre of the frame. On a single presenter that is nearly invisible: somebody who stands still is framed almost identically whether the tracker is running or not, which is precisely why it survived as long as it did before anyone caught it. On a two-person split the same failure is obvious inside a second, because the panes stop following anybody and one of them settles on the wall over a shoulder. If you ever suspect tracking in your own output, test it on multi-speaker footage first — it breaks loudly there and quietly everywhere else.
A 4-up on a 1080 × 1920 output hands each tile 540 pixels of width, and that number governs the result more than the choice of layout does. Close-framed webcam footage, where a head already fills its source frame, scales down into 540 and stays readable. A single camera at the back of a room gives each face a small patch of the original frame and then asks for that patch to be enlarged up to 540, and enlargement does not restore detail that was never recorded in the first place. Measuring how tall a head actually is in your source predicts the quality of a grid better than anything in the settings.
Manually this means duplicating the clip, masking each speaker, positioning both, and keyframing the crops whenever someone shifts in their chair. For a single hero clip that is a reasonable afternoon. For every clip from a weekly show it is a job, and it is the kind of job that gets skipped first when a deadline moves.
A proper multicam edit with ISO recordings gives you the most control there is, including full-resolution close-ups of each speaker. It also requires the recordings, the sync, and someone who knows the workflow. This works from the single video you already published, which is why it fits people who do not have an editor.
Auto-reframe features in general editors follow motion or a single subject, which handles a solo talking head perfectly well. Point one at a panel and it either sits between people or oscillates between them. The layouts here exist specifically because a one-subject assumption breaks on a conversation.
Split screen is fairly widely offered now; 3-up and 4-up are less so, and the quality gap is usually visible in whether the layout adapts when the number of visible faces changes mid-clip. Run a panel episode through and look for empty panes and reshuffling tiles — those are the tells.
Take an episode where the reaction shot matters as much as the answer. Watching the split version next to a centre-cropped one settles the argument in about ten seconds.
⚡ Clip My Conversation — $1 trial
