Lectures are recorded in parts, in a studio, with someone operating the camera. Before the lecturer says a word to the students, they read a slate so the editor can tell the parts apart. When they move to a new slide they often say so out loud, so whoever builds the slides knows where it goes. And when a sentence comes out wrong, they stop and say it again.
All three are good studio practice. None of them belongs in the version a student watches. And all three are easy to miss, because each one sounds like ordinary speech.
Teachers watching finished lectures caught the moments that had slipped through, and we changed the engine so it listens for each kind on its own. The clips below come from one undergraduate database lecture. On the left is the raw camera file, exactly as it was recorded. On the right is the finished lecture at the same moment, with the rest of the pipeline — slides, captions, teaching visuals — built on top.
1. The slate
The first thing on the recording is the lecturer reading out the semester, the unit and the script number. Then comes the real opening.
“… unit five, script eight.” “Hello and welcome back to …”
“Hello and welcome back to …”
A slate is easy to spot when it stands alone. The harder cases are a slate run straight into the welcome in one breath, or a slate in the middle of the file because a new camera part starts there. The engine looks for the slate's shape — a unit, a number, then a script or part number — wherever it occurs, not only in the first few seconds.
2. “Slide number two.”
This lecture says “slide number …” out loud eleven times, once at every change of slide. In the studio it is a note for the slide builder. In the finished lecture the slide itself changes at that moment, so the spoken note is just noise in the middle of a thought.
“… teaches you to use them correctly. Slide number two. In this session, here is our path …”
“… teaches you to use them correctly. In this session, here is our path …”
Removing them had a knock-on effect. The writers who build each lecture's slides had often pinned a slide to exactly those words, because they were the clearest marker in the transcript, and the cut broke the pin. The engine now moves each pin to the next words actually spoken, and checks that every slide still lands within a second of where it was meant to.
3. The restarted sentence
Six minutes in, the lecturer starts a sentence about a bill that must reconcile to the last paisa, stumbles, and starts again. Both attempts are in the recording, one after the other.
“… must be reconciled exactly to the paisa, like our canteen. This is the … must be reconciled exactly to the paisa, like our canteen bill …”
“… must be reconciled exactly to the paisa, like our canteen bill …”
This is the one that needs judgement. Lecturers repeat themselves on purpose all the time: for emphasis, for a student who missed it, to recap. A repeated phrase alone is not a reason to cut. The engine treats it as a restart only when the first attempt trails off unfinished and the second picks up the same words and carries on. It always keeps the later attempt, because that is the one the lecturer chose.
How the engine decides
Transcribe every word with its time. The whole check runs on the transcript, so a cut always lands between two words and never inside one.
Look for each kind of studio talk. Slates, spoken slide numbers, remarks to the crew (“okay, let's go”, “rolling”), “repeat” and “retake” markers with the line said twice, abandoned restarts, and anything said before the welcome.
Check the order. A lecture opens with its welcome. If the welcome turns up minutes in, the camera parts were probably joined in the wrong order, or part of another lecture got attached. Cutting cannot fix that, so the lecture is held for a person instead.
Refuse to over-cut. If the cuts would remove more than about an eighth of what was said, something unusual is going on. The lecture goes to a person rather than being trimmed on a guess.
Listen to the finished lecture. Every other check reads the recording. This one transcribes the finished lecture — what a student will actually hear. If a slate, a slide callout, a crew remark or a second welcome is still audible, the lecture is not delivered.
We would keep that last step even if the others were perfect. A cut can be planned correctly and still be lost later in the pipeline. Checking the finished lecture by ear catches that, whatever caused it.
What it came to on this lecture
| Removed | Count | Why it was in the recording |
|---|---|---|
| Slate before the welcome | 1 | Labels the camera part for the editor |
| “Slide number …” read aloud | 11 | A note for whoever builds the slides |
| Restarted sentence | 1 | The first attempt came out wrong |
| Long pauses, shortened to half a second | 9 | Thinking time and slide changes in the studio |
| Length | 17:23 | 16:22 — a minute shorter, nothing taught removed |
None of these moments is the lecturer's mistake. They are what a well-run studio recording sounds like. Taking them out is an editor's job, and the engine now does it on every lecture.
The check runs on every new lecture, and we have re-run it over lectures already delivered. Anything it flags is rebuilt and replaced at the same link, so nothing a student has bookmarked stops working. For the wider cleanup pass this belongs to, see the two passes nobody asks for.
Send us a raw recording — slates, retakes and all — and we will send back the cleaned lecture with a list of every cut and the reason for it. Send a lecture →