Applause in a Lecture Recording: Where the Words Go

Two places a clap can land

A lecture recording makes two different noises. The room is noisy between sentences - coughs, chairs, someone's phone - and sometimes the room is noisy while you are talking. Those are not the same problem and they do not have the same answer.

The test source is a 13.46-second read of five sentences, 44.1 kHz mono, peaking at -1 dBFS. It returns 248 characters and 50 words, and every file below is measured against that.

• Control: 13.46 s, 1,187,604 bytes, 248 characters, 50 words.
• Applause in a gap: 0.5, 1, 2, 3 and 5 seconds, spliced at the 3-second mark, between sentences.
• Applause over a sentence: 3 seconds starting at 2 seconds in, laid on top of the speaker at -6, 0, +6, +12 and +18 dB.
• Applause tagged onto the end: 3 seconds after the last word.
• The clap is band-limited noise with a density envelope, not flat white noise.

Five seconds in a gap changes nothing

Every file that had the clap between sentences returned 248 characters and 50 words. Not a word moved, not a comma moved. The file size grew with the duration - 1,231,704 bytes at half a second up to 1,628,604 at five seconds - and the text did not change at all.

Five seconds is a long time. It is longer than the six-second silence threshold that breaks a transcript, which is the obvious thing to worry about, and it did not fire here. The difference is that this audio is loud. A gap full of noise is not a gap as far as the model is concerned.

That is the single most useful result in this test for anyone recording a talk. If the applause lands after your sentence, leave it. Trimming it costs you work and buys nothing.

Applause over the last word

Three seconds of applause after the last word returned 248 characters and 50 words as well. The transcript simply ended where the speech ended. Nothing was invented after it, and nothing was attributed to it.

That contradicts what a long silence does. Six seconds of pure silence after the same speech produced 2,077 characters in an earlier test on this site, with the real transcript followed by repeated filler. Noise does not trigger that. Silence does. If your recording ends with a hiss rather than a digital zero, you are safer than you think.

Applause over the speaker: 6 dB is the line

Now the other case. We laid three seconds of the same clap starting at the two-second mark, right on top of the speaker, and raised it in steps. At 6 dB below the voice, all 50 words came back. At equal loudness, 49 of 50 matched and one phrase changed. Above that, the losses accumulate.

• -6 dB under the voice: 248 characters, 50 words, 50 of 50 matched.
• Equal loudness: 245 characters, 50 words, 49 of 50 - “before the meeting” became “before the news”.
• +6 dB over the voice: 237 characters, 49 words, 45 of 50.
• +12 dB over: 229 characters, 48 words, 43 of 50.
• +18 dB over: 208 characters, 43 words, 41 of 50.

Read the character counts down that list. They fall steadily: 248, 245, 237, 229, 208. Words disappear and the text gets shorter, which is the opposite of what noise at a moderate level does. In the earlier noise test the text got longer as it got worse, because the model started writing sentences. Here it just loses material.

That difference matters when you are diagnosing your own file. A transcript that is shorter than the audio should produce means something covered the voice. A transcript that is much longer means something confused it.

What actually gets lost

At +6 dB the transcript still reads like a lecture. The losses are inside phrases: “out loud before the meeting” became “out loud for you”, and “the team agreed to review” became “the team agrees to remove”. Those are the same words in roughly the same shape, which makes them easy to miss when you skim.

By +18 dB whole clauses collapse. “We are going to read the whole report out loud” came back as “We are going to read the whole report out last year”. The second sentence lost its opening entirely. That is a transcript you would have to redo, not edit.

Why the gap is safe and the overlap is not

Between sentences there is nothing to compete with. The model has already decided the previous sentence ended, and a burst of noise in that space does not give it anything to transcribe. It has no reason to start a new sentence there, and it does not.

On top of a sentence, the clap competes for the same time slice as the voice. A voice and a broadband burst occupy the same frequencies, and the louder one wins the slice. At +6 dB the clap is winning about half of them, and the words that need those slices go.

What to do with a lecture recording

The practical rule is about position, not level. Applause at the end of a sentence is not worth editing. Applause or crowd noise while someone is talking is, and no amount of gain riding fixes it because you would be turning the voice down too.

• If the noise sits in the pauses, upload as is. Five seconds in a gap cost zero words.
• If the noise sits under the voice, split the file at that stretch and upload the clean part.
• Do not normalise the whole file to fix a noisy passage. That lifts the noise with the voice.
• If the transcript is shorter than you expected, look for a passage where something covered the speaker.
• If the transcript is longer, the problem is silence, not noise.

The one thing this test does not cover

Real applause is not one clap. It is dozens of hands at slightly different times, which makes it denser and less predictable than the envelope we built here. That would push the +6 dB result earlier, though the shape of the curve would hold: a flat region, then a steady decline.

The 50-word source is also short. A five-minute recording has more chances to land on a bad slice, so a per-file word loss rate from this test will look optimistic against a real lecture. What carries over is the direction: noise in a pause is free, noise on a word is not.