Why AI clips cut off mid-sentence, and how we measured it

If you've used any AI clipping tool for more than a week, you've seen this: a genuinely good 40-second moment that ends on "...and the part everyone misses is", and stops. The clip is unusable. Not bad, unusable, because the payoff was two seconds past the cut.

This isn't the model having poor taste. It's a specific, boring technical limitation, and it's worth explaining because it tells you something about how these tools are built.

The root cause: transcripts without punctuation

To cut a clip, a tool needs to know two timestamps: where to start and where to stop. To choose those intelligently, it needs a transcript with word-level timing, which word was spoken at which second.

Here's the catch. When you ask a speech-to-text API for word-level timestamps, several of them return no punctuation at all. You get this:

so in college i was a government major which means i had to write
a lot of papers now when a normal student writes a paper

Every word has a start and end time. Not a single full stop.

So the tool knows exactly when each word was said, and has no idea where any sentence begins or ends. Choosing a clip boundary from that is guesswork.

What tools do instead: guess from pauses

The standard workaround is to infer sentence boundaries from silence. If there's a gap of, say, 0.45 seconds between two words, call it a boundary. A longer gap, maybe 0.8 seconds, call it a strong boundary. Probably a full stop.

It's a reasonable heuristic and it half-works. It also fails constantly, because:

• People pause mid-sentence, for emphasis, all the time.
• People run sentences together with no pause at all when they're excited, which is exactly when the good moments happen.
• A pause after "and" or "but" is someone thinking, not finishing.

You can patch around the last one with a list of words that shouldn't end a clip. We did. It still isn't good enough, because the underlying signal just isn't sentence structure, it's breathing.

Measuring it instead of assuming

The useful step is to stop reasoning about it and count. We took a real production job: a 14-minute talk, 10 clips generated, and checked every boundary against the transcript, asking two questions: does the clip start at the beginning of a sentence, and does it end at the end of one?

The result:

	Result
Clips starting mid-sentence	7 of 10
Clips ending mid-sentence	7 of 10
Overlapping clips (near-duplicates)	3 pairs

Seven out of ten. On clips the system was confident enough to ship. That matched the complaint exactly, and it turned a vague "the cuts feel off" into a number we could try to move.

The fix: get punctuation back

The reason we could fix this is that we'd changed transcription engines for an unrelated reason, cost. Running speech-to-text locally instead of through a paid API removed a per-minute charge and, as a side effect, the local engine returns real punctuation in its word tokens:

[0.3-4.0]  So in college, I was a government major.
[4.0-8.2]  Which means I had to write a lot of papers.

Now the sentence boundaries are simply there. No inference required.

Two things changed off the back of that:

1. The model gets a sentence-indexed transcript. Instead of a wall of words with a timestamp every five seconds, where it has to interpolate a position and inevitably lands mid-phrase, it receives one line per sentence, with that sentence's exact start and end time, and is told to reuse those numbers verbatim. It can no longer land mid-sentence by accident, because it's picking from a list of sentences.

2. Boundaries snap to punctuation, with pauses as fallback. The old pause heuristic still exists, because music, crosstalk and heavy accents still produce stretches with no punctuation. But it's now the backup, not the primary signal.

We also stopped trusting the model on overlaps and enforced it in code: clips are de-duplicated after boundary-snapping, keeping the higher-scoring one when two collide.

The result, measured the same way

Same video, same measurement:

	Before	After
Clips starting mid-sentence	7 of 10	0 of 7
Clips ending mid-sentence	7 of 10	0 of 7
Overlapping clips	3 pairs	0

Every clip now opens and closes on a complete sentence. In practice that looks like a clip that starts "But then came my 90-page senior thesis." and ends "...I wrote 90 pages over 72 hours, pulling not one, but two all-nighters.": a complete thought with a beginning and a payoff.

Note the clip count dropped from 10 to 7. That's the de-duplication removing near-identical overlapping segments. Fewer, better clips is the right trade.

What this doesn't fix

Punctuation solves where to cut. It doesn't solve what's worth cutting. A clip can be perfectly bounded and still boring, that's a judgement problem, and it's why every clip we produce also carries a score with its reasoning shown, so you can disagree with it.

It also degrades with audio quality. Heavy overlapping speech produces worse transcription, worse punctuation, and therefore worse boundaries. Clean audio in, clean cuts out.

Why this matters when you're choosing a tool

You can test this yourself in about five minutes, on any tool: generate some clips and check whether each one starts and ends on a complete sentence. Not whether it feels right, literally read the first and last words.

If a meaningful share of them are cut mid-thought, you'll be trimming every clip by hand, and the automation isn't saving you the time it claims to.

If you want the pricing-model comparison rather than the quality one, see GetClipMachine vs Opus Clip. If you're clipping long-form conversation specifically, AI clipping for podcasters covers the workflow.