The AI clipper that does not cut mid-sentence
Picking the moment and placing the cut are different problems. Most tools solve the first and let a clock solve the second, which is why clips open on a pronoun with nothing to attach it to and end halfway through the punchline. Here is exactly what Momevera does instead.
Free tier, no card, 90 credits a month. Free exports are 720p and carry no watermark; paid plans raise the resolution.
The problem, stated precisely
An AI clipper has to answer two questions, and they are not the same question. Where is the moment is a scoring problem — loudness, motion, shot changes, speech density, laughter. Where does the clip that carries it start and stop is an editing problem, and it is answered from the transcript, not from the score.
Conflating them is the standard failure. Score a window, take a fixed length around the peak, clamp it to the length the user asked for, ship it. Every one of those steps can move a boundary, and none of them can see a sentence. The result is the complaint everyone in this category has heard: the clip opens on and then he just, and it ends one word before the payoff.
In Momevera the length clamp cannot override the sentence snap, because the clamp is expressed in sentences. A clip is lengthened or shortened by moving to a different sentence boundary. When no pair of sentence boundaries inside a candidate satisfies both the floor and the ceiling, the candidate is discarded and the next one is scored. This is the reverse of the ordering that produces mid-word cuts.
The four rules, with their actual numbers
Start — walk back to the setup, never into the previous scene
If the sentence the clip would open on points at something said before it — an opener like and then he, that is why, which is, or a bare pronoun with no referent inside the clip — the start walks backwards one sentence at a time until the opening line stands on its own. Two hard limits stop that walk running away: it never crosses a silence of 1.2 seconds or more, because that is a scene change rather than a sentence break, and it never adds more than 12 seconds. Leading discourse markers are then trimmed (so, well, um, okay — at most two, and never below five remaining words), and the cut is placed 0.12 seconds before the first word so the opening consonant is not clipped.
End — land after the payoff, and keep the reaction
The clip ends at the sentence end at or after the payoff. If a laugh lands within 2 seconds of that, the laugh is included and the cut falls 0.5 seconds after it. If instead the next two seconds are wordless and quiet, the cut falls 0.5 seconds after the last word rather than sitting in dead air. A reaction is caught two ways. One is a written-out laugh in the transcript, matched deliberately narrowly, so ha on its own is a word and haha is a laugh. The other needs no transcript at all: a second of audio at or above a loudness threshold carrying at most one spoken word, which is what an unlabelled laugh, a gasp or a room reacting actually looks like in the energy curve.
Gate — a question is a setup, not an ending
A final sentence that is a question, or that ends on a conjunction, is a setup. The clip is extended forward a sentence at a time, up to three, until it lands on one that actually ends; if none does, the candidate is dropped. The conjunction test has two tiers: a strong list (and, but, so, because, which, or, if, when, while, though, although, however, plus) that fails regardless, and a weak list (that, then, with, for, to) that only fails when the transcript did not itself mark the sentence as terminated — so a line that genuinely ends on for. is allowed to.
Clamp — length is enforced in sentences
Only now is the requested length applied, and it is applied by choosing a different sentence boundary. If you set a hard band, it is satisfied by moving to a different sentence pair rather than to an arbitrary offset, and on any source long enough to hold your floor nothing delivers outside it. The one exception is a source SHORTER than the floor you asked for: there the floor collapses to the length of the video rather than returning nothing at all. That matters because short-form monetisation is gated on duration, so a creator who needs every clip over a minute needs a floor that actually holds.
What a sentence is here
A sentence is a run of words the transcript punctuated, paused between, or ran long between. That third clause is doing real work: platform auto-caption tracks frequently arrive with no punctuation at all, and a resolver that only trusts full stops would see one sentence three hours long.
- Ends a sentence
- A terminal . ! or ? on the word.
- Also ends it
- A measured gap of 0.5 s or more before the next word.
- Backstop
- A 30-word cap, for transcripts that give neither.
- Counts as silence
- 0.35 s of real silence between two spoken words — measured after subtracting the padding Whisper smears into a gap (up to 0.16 s a side), so roughly a 0.67 s raw gap — merged with any range the transcriber flagged as speechless.
One segmentation, one set of options, three callers. The boundary resolver, the candidate enumerator, and the code that quotes a delivered clip's opening and closing line in the editor all import the same options object rather than each carrying its own numbers — three literals drifting apart would put a quote on screen that the picker never used.
You can read the cut it chose
The most deliberate thing in the pipeline used to read as luck, which is the same failure as not doing it. So every clip now shows its own first and last sentence, as spoken, in the clips list and in the editor.
It is derived at display time from the clip's stored word timings, using the same segmentation function the resolver used — not stored on the row. That is deliberate: a stored field would be blank on every clip rendered before the feature existed, and a blank quote says the opposite of what is true. Trims you make afterwards are not folded in, because those lines describe the cut the picker chose, and the editor says so beside them.
What this does not do
Sentence-boundary cutting is not a promise that a clip will perform. It is a promise about one specific defect. Some things it will not fix:
- A source with nothing in it. The picker scores against an absolute quality bar, so a flat source returns fewer clips rather than a full batch of weak ones — see the guide on long streams.
- A transcript that is wrong. If the transcriber misheard the word, the sentence boundary is still in the right place and the caption still says the wrong thing.
- Music and wordless footage. Two of the four evidence signals — speech density and transcript events — need a transcript; measured loudness and visual motion do not. Without one the bar is lower and the boundary rules have less to work with.
- Your editorial taste. The resolver stops clips ending mid-thought. It cannot know that the third joke was the one worth posting.
If you want the fuller picture of how a clip is scored before any of this runs, the long-stream guide covers the quality floor and why a three-hour VOD can honestly return seven clips instead of twelve.
Questions, answered
Why do AI clippers cut mid-sentence in the first place?
Because the usual pipeline picks a moment and then takes a fixed window around it — say 30 seconds centred on an audio peak. The window boundary lands wherever the clock says, and the clock knows nothing about speech. A length clamp applied after the moment is chosen will always be able to override a word boundary, which is exactly the ordering Momevera reverses: the length band is satisfied by choosing a different sentence pair, never by moving the cut to an arbitrary offset. If no sentence pair inside a candidate fits the band, that candidate is dropped and the next one is tried.
What happens if a clip would end on a question?
It is treated as a setup, not a payoff, and the clip is extended forward a sentence at a time, up to three, until it lands on one that actually ends — or dropped if no extension fits the length band. The same rule catches an ending that trails off on a conjunction: and, but, so, because, which, or, if, when, while, though, although, however, plus. A clip that ends mid-thought is worse than a clip that does not ship.
Does it cut before the laugh?
No. If a reaction lands within 2 seconds of the payoff sentence ending, the reaction is included and the cut falls about half a second after it. Cutting before the laugh throws away the thing the clip was chosen for. If the following two seconds are wordless and quiet instead, the cut falls half a second after the last word rather than sitting in dead air.
Can I still see where the cut landed?
Yes. The clips list and the clip editor both quote the clip's opening and closing sentence, derived from that clip's own stored word timings using the same segmentation the boundary resolver used. It is re-derived at display time rather than stored, so it reads correctly on clips rendered before the feature existed.
Does this work on a transcript with no punctuation?
Yes. Platform auto-caption tracks routinely arrive with no punctuation at all, so a sentence is also ended by a measured pause of half a second between two words, with a 30-word cap as a backstop. The pause is measured from the word timings, not guessed.
What if I want to move the cut myself?
Trim it in the editor. The resolver decides where the clip starts and stops before it renders; anything you trim afterwards is yours, and the editor labels the quoted opening and closing lines as the cut the picker chose so the two are never confused.