Two-person footage breaks the two things automated clipping usually gets wrong. A centre crop frames the empty air between two chairs, and an answer lifted away from the question it responds to stops making sense. This cuts around the exchange, follows whichever person is speaking, and hands back vertical clips where the setup and the payoff are both still there.
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
The interesting work happens in step two, and none of it is yours.
Paste the link if the interview is already on YouTube, or upload the recording straight from your machine. Remote sessions exported from a call recorder work, and so does a two-camera edit that has already been cut together. Two hours is the ceiling on a single paid upload; the free demo takes anything shorter than thirty minutes.
The transcript is read as a dialogue rather than a wall of speech. It tracks who is asking and who is answering, which questions produced a real answer instead of a deflection, and where an exchange resolves. That structure is what lets a clip start on a question and finish on the beat after the point lands, rather than fifteen seconds into an answer nobody has context for.
Each clip comes back framed, captioned and scored 0-100. Watch the top few, extend a boundary if you want the interviewer's reaction included, and publish. Send a copy to your guest as well — a good interview clip is the one piece of content a guest will reliably reshare, and their audience is the reason you booked them.
These are specific to two people talking. On a monologue most of them are irrelevant, which is precisely why generic tools do not bother with them.
When the moment depends on both people — a challenge and a rebuttal, a claim and a raised eyebrow — the split layout stacks them so neither is cropped out. The alternative that other tools produce is a clip where you hear a question from someone who is not on screen, which reads as an editing mistake even to a viewer who cannot say why.
A fixed crop on a two-shot spends half the clip on whichever chair was closer to the middle. Speaker tracking moves the frame to whoever currently has the floor, so a fast back-and-forth stays legible in vertical instead of turning into a shot of the space between two people.
An answer is often meaningless without the ten words that prompted it. Boundaries are chosen with the exchange in mind, so a clip can open on the question and close after the answer completes. That single behaviour is what separates an interview clip that works from a fragment that needs a caption to explain itself.
Nobody in conversation speaks in tidy blocks, and the temptation for automated tools is to slice on silence. Silence in an interview is usually a person thinking, not a person finishing. Cutting on sentence boundaries respects the difference, so a clip does not begin halfway through a considered thought.
A long interview contains many decent exchanges and two or three genuinely strong ones, and after two hours in the room you are the worst-placed person to tell them apart. The grade sorts them for you on how the clip opens, paces and peaks, which is a more useful second opinion than asking a colleague to watch the whole thing.
Interviews are the format most likely to be watched on mute in a feed, and the format where a strong regional accent or an unfamiliar name most needs to be on screen. Word-by-word captions in eleven styles are rendered into the video, timed to the speech, and they stay correct if you trim the clip afterwards.
Conversational footage carries more of this than any other format — the half-started question, the overlapping agreement, the pause while somebody remembers a name. Stripping it automatically usually buys back several seconds on a minute-long clip and makes both parties sound sharper than the raw recording did.
Export vertical for Shorts, Reels and TikTok, square for feeds, or landscape if the guest wants something for a newsletter. Handing a guest three finished clips within a day of recording is the cheapest way to make them promote the episode, and it costs you nothing extra once the clips exist.
If the conversation is being streamed to YouTube, Twitch or Kick, clips are produced during the broadcast rather than after. A remark that will be quoted everywhere tomorrow can be posted while people are still watching, which is not something a post-production workflow can offer.
The whole format is one guest and one host for ninety minutes. Clipping it is the growth channel, and doing it manually is why most shows stop after four episodes. See the podcast clip generator.
Long-form conversations get discovered through short excerpts of the sharpest exchange. Framing decides whether that excerpt is watchable on a phone.
A recorded sit-down usually contains one quotable admission, and the value of getting it out fast decays by the hour rather than by the day.
Vox pops produce dozens of short exchanges in a single shoot. Sorting them by score is faster than reviewing four hours of footage looking for the three good responses.
Employee interviews and culture pieces work far better as sixty-second answers than as a four-minute corporate film nobody finishes.
A customer answering why they switched is the most persuasive footage a company owns. The webinar clip generator covers the B2B side of the same idea.
Shoot an interview properly and you get a wide two-shot: both people, some room around them, a comfortable amount of air. It looks great on a television and it is close to unusable on a phone. Converting it to vertical removes about two thirds of the width, and if that crop is fixed to the centre it will often frame the coffee table, the microphone stand, or a nicely lit patch of wall.
Even a crop that finds a face has a second problem. Interviews alternate, so a crop locked on the host misses every reaction from the guest and vice versa. The viewer hears a voice with no source, which is subtly disorienting in a way most people register as the clip feeling cheap.
The fix is not clever, it is just work: identify who is talking, move the frame accordingly, and when both people matter simultaneously, stack them rather than choosing. Doing that by hand on every clip is a tedious afternoon. Doing it automatically is the reason this page exists.
Take a strong answer out of an interview and you often find it starts with a pronoun. "That happened to us twice." Twice what, and to whom? The line was perfectly clear in the room, because the question was still hanging in the air, and it is nonsense fifteen seconds later in a feed.
There are two ways to solve it. You can write an on-screen caption explaining the setup, which is what most people do and which costs you the first two seconds of attention. Or the clip can simply include the question. The second is almost always better, because a question is short, it creates tension, and it gives the viewer a reason to stay for the answer.
This does not always work. Some interviewers ask forty-second questions with three sub-clauses, and no reasonable clip can carry that as an opening. In those cases the clip starts inside the answer at the first line that stands alone, and you may want to add your own context. Knowing which situation you are in is a ten-second judgement once the clip is in front of you, and impossible before it exists.
Heavy crosstalk is the hardest case. Two people talking over each other produces a transcript with uncertain speaker boundaries, and uncertain boundaries produce cuts in slightly wrong places. Panel discussions where three people interject are worse than a calm two-hander by a wide margin.
The second failure is a guest who cannot be brief. Some people answer in four-minute paragraphs where the point arrives at the end and depends on everything before it. There is no short clip inside that answer, and no tool can invent one. If your interviews consistently run that way, the fix is upstream — asking sharper questions, or interrupting more.
Third, audio recorded on a single room mic with both people some distance away transcribes poorly, and a poor transcript damages captions and cut points together. A cheap lapel mic on each person improves clip quality more than any software choice you will make this year.
Almost everything that determines clip quality is decided before the camera stops. Four habits are worth building, and none of them cost anything.
Ask short questions. A twelve-word question can open a clip; a ninety-word preamble with three sub-clauses cannot, and it forces every clip from that answer to begin cold. If you catch yourself explaining the question, stop and ask it again in one sentence — you can always cut the first attempt.
Let a beat of silence sit after an answer finishes. Interviewers instinctively fill the gap with agreement noises, and those noises land exactly where the clip wants to end. A held pause gives the clip a clean edge and often prompts the guest to add the better line.
Encourage the guest to restate the subject rather than lean on a pronoun. "The migration took us eight months" survives being lifted out of context; "that took us eight months" does not. Say so before you start recording and most people adapt within ten minutes.
Mic both people separately and frame slightly tighter than feels right for a landscape edit. A generous wide two-shot looks handsome on a monitor and leaves the vertical crop working with a fraction of the pixels. Everything downstream — the tracking, the captions, the boundaries — is downstream of these choices.
The part people price is the cutting, and the cutting is nearly free. Everything either side of it is where the hours go. Say you have an hour-long conversation and you want eight clips out of it: that is the better part of an hour scrubbing to find the exchanges, eight separate decisions about where each one opens, and then the reframing — which on two-person footage is not one crop but a sequence of them, one per speaker change, keyframed across the length of the clip. That last part is why editors quietly resent vertical interview work.
A human still wins on two things and it is worth naming them precisely. They know the guest, so they know which line will get quoted approvingly and which will get screenshotted uncharitably. And they hear tone. A sarcastic answer lifted out of its exchange reads as sincere in a feed, and no transcript carries the difference.
What a person loses is every clip after roughly the sixth. Reframing is mechanical, repetitive and identical each time, and attention on repetitive work degrades in a way that surfaces as inconsistent framing across a week of posts. Keeping the judgement and handing over the mechanics is not a compromise between the two approaches — it is the division of labour that matches what each side is genuinely good at.
The honest test is throughput. If you publish two interview clips a month, cut them yourself and enjoy the craft of it. If the plan calls for a clip every weekday from a weekly episode, that plan does not survive contact with manual reframing, and most shows that attempt it quietly stop by the second month.
Platform clip buttons are the free option and they beat nothing at all. YouTube has its own clip feature and Twitch and Kick both ship one, and if a moment is genuinely good your audience will mark it without being asked. What comes back is a landscape excerpt with no captions and no vertical framing, so treat it as a signal about which moments landed rather than as a finished post.
A transcript editor is the second route. Deleting a paragraph of text deletes the matching stretch of video, which suits anyone already reading the transcript to write show notes. It removes the scrubbing but not the choosing — every clip is still your decision — and it does nothing whatsoever about framing a two-shot for a phone.
The overlooked one is the guest. Anyone with an audience of their own usually has somebody who edits their content, and a raw file plus three suggested timestamps frequently comes back as three finished clips at no cost to you, cut by a person motivated to make their principal look good. Most podcasters never ask, which is odd given the guest's editor is already on a retainer and the guest wants the episode to travel as much as you do.
Then there is the null option: publish the full interview with chapter timestamps and do nothing else. It is a real strategy and it fails for a specific reason. Timestamps are navigation for someone who already opened the video; clips are discovery for someone who never would. They are not substitutes, which is why a show that only does the first tends to grow strictly within the audience it already has.
Observations from operating the pipeline in production — not general advice.
There is no speaker-diarisation step producing neatly labelled turns — the words arrive as one continuous transcript. Turn-taking is inferred from the shape of the language instead: a question, an answer that resolves it, a name being addressed directly. That holds up on a calm two-hander and breaks on sustained overlap, where two voices collapse into a single run-on sentence and the cut lands somewhere inside it. It is the mechanical reason panels need more review than a sit-down, and why separate microphones improve the transcript before they improve the sound.
People assume raw wide footage is the safe thing to hand over and that a finished multi-camera edit will confuse the framing. It runs the other way round. A cut between an A-cam and a B-cam gives the tracker one face per shot with a hard boundary between them, so each stretch gets its own anchor and nothing has to arbitrate between two equally valid subjects. The genuinely difficult input is a single locked-off wide two-shot held for ninety minutes, because every frame contains two candidates and the choice has to be remade continuously.
Across every format we see, the adjustment made most often on interview output is dragging the opening back a beat — not to repair a bad boundary, but to catch the interviewer's reaction or the scrap of setup that makes an answer land. Boundary logic optimises for a clean start on a complete thought, and a reaction is not a thought, which is why it sits just outside the cut by design. So a clip that starts correctly and still feels abrupt usually wants two more seconds at the front rather than a different moment.
Manually, the cutting is the easy part. The cost is rewatching ninety minutes to find the six exchanges worth keeping, then reframing each one so both people stay visible. Do that weekly and it becomes the reason the show stops publishing clips. Reviewing a scored shortlist takes twenty minutes and produces more posts than the diligent version ever did.
Most of them are tuned for one person addressing a camera, and they will confidently centre-crop your two-shot. On a solo video you would not notice the difference. On an interview it is the whole ballgame, and the tell is a clip where a disembodied voice asks something while the wrong person sits silently on screen.
A good editor will beat this on the pieces that matter most, because they understand your show, your guest and what your audience finds funny. They will not beat it on cost per clip or on turnaround, and turnaround is what determines whether a clip goes out while the episode is still current. Use a person for the trailer and the launch clip, and automate the other fifteen.
A long interview is discovered by people who already follow one of the two people in it. Clips reach everyone else, and they are also what convinces a guest's audience that the full conversation is worth ninety minutes. Both belong in the plan; only one of them is a distribution channel. Try it on your last interview and compare what the score picked against what you would have chosen.
You already know which exchange was the best one in that interview. Check whether the score agrees, and look at how the framing handled your two-shot.
⚡ Get A.I Clips — $1 trial
