Distorted Audio: What Clipping Does to a Transcript

One clip, six levels of clipping

The test file is a spoken list of five sentences built around sibilants, because fricatives are the first thing a clipped recording loses. It runs 16.94 seconds, 1,494,540 bytes, 44.1 kHz mono, 16-bit PCM. The control copy sits at -1 dBFS gain and peaks at -4.17 dBFS, which is a normal healthy recording. It returns 285 characters and 50 words.

• Control: 16.94 s, 1,494,540 bytes, peak -4.17 dBFS, 285 characters, 50 words.
• Hard clipping: raise gain by 3, 6, 12, 24, 36, 48 and 60 dB, then clamp every sample at plus or minus full scale.
• Soft limiting: raise gain by 6, 12 and 24 dB through a limiter instead of a clamp.
• Every file uploaded once, one at a time, same endpoint, same day.
• Word count compared against the control with an edit-distance pass, not a raw position match.

29% of samples flat, one capital letter moved

At 24 dB of gain the peak is pinned at 0.000265 dBFS and 29.15% of all samples sit at full scale. That is a file that sounds like it is tearing. It came back 285 characters and 50 words, exactly the control length, with 49 of 50 words untouched. The single difference was the word “shipment” coming back capitalized mid-sentence. That is not a word error. It is a formatting quirk.

The same thing happened at 12 dB of gain, where 1.93% of samples were flat. Identical 285 characters, identical 50 words, identical capital letter. A quarter of the waveform gone and the text does not notice.

At 58%, the ordinal goes numeric

The 36 dB file is where a real change shows up. 58.2% of samples are flat, the peak is pinned, and the output is 281 characters instead of 285. The word count holds at 50. Two words changed, and they are the same word twice: “Sixth” came back as “6th”.

• 58.2% clipped, 281 characters, 50 words, 48 of 50 matched.
• Change one: “the shop on Sixth Street” became “the shop on 6th Street”.
• Change two: “the Sixth shipment” became “the 6th shipment”.
• Nothing else in five sentences moved.

That is worth sitting with. The model did not invent a sentence, drop a clause or lose the thread. It wrote the same address with a numeral instead of a word. You can still read it and still search it.

74% and 82%: an extra article, nothing else

Push to 48 dB and 74.11% of the file is flat. Push to 60 dB and it is 82.33%. Both returned 285 characters and 51 words, and both matched 48 of 50. The extra word is a “the” that appears in front of “6th Street”. The rest of the transcript is word for word what the clean file gave.

So the shape of the damage is flat, not steep. You lose a handful of specific tokens, mostly numbers and ordinals, and the model keeps reconstructing the rest from context. Five sentences of ordinary speech survived four fifths of their waveform being erased.

A soft limiter does not cost a single word

Running the same gain through a limiter instead of a clamp is cheaper than you would expect. At 6 dB and 12 dB the transcript came back 50 of 50, zero edits, 285 characters. At 24 dB it returned 49 of 50 with the same capital letter the hard clamp produced.

That matters for how you record, not just how you repair. A limiter rounds the peak instead of slicing it, and the model reads the rounded version as the same phoneme. A clamp removes the top of the waveform and leaves a flat shelf where a vowel shape used to be. The shelf is what starts to look like a different number.

Turning it back down does not undo it

We took the 29% clipped file and pulled it back down 25 dB so it peaked below full scale again. It returned the same text as the clipped original, capital letter and all. You cannot un-clip a file. The samples that were flattened are gone, and lowering the level just makes quiet flat samples.

The good news is that the text was already fine at that level of damage. The repair you would want is not on the audio at all. It is reading the output and fixing “6th” back to “Sixth” if your house style wants the word.

Re-encoding afterwards changes nothing

Two more uploads to close the loop. The 58% clipped file re-encoded to 128 kbps MP3 returned the same 281 characters, the same 50 words and the same two numerals. The clean control re-encoded to 128 kbps MP3 returned 50 of 50 with zero edits.

• Clean WAV: 285 characters, 50 words, 50 matched.
• Clean WAV to MP3 128k: 285 characters, 50 words, 50 matched.
• 58% clipped WAV: 281 characters, 50 words, 48 matched.
• 58% clipped WAV to MP3 128k: 281 characters, 50 words, 48 matched.

Whatever the clipping did, the codec did not add to it. That is the opposite of what most people assume when a transcript looks wrong and the file sounds bad. The loss happened at the ceiling, not in the encoder.

Why the model holds up better than your ears

Speech recognition does not read a waveform the way you do. It converts short slices into a description of frequency content and then matches patterns across time. A clipped vowel keeps its formant peaks even when its top and bottom are shaved off. The peaks are what carry the identity of the sound.

Fricatives are the exception, and that is why “Sixth” broke first. The “s” and the “th” live almost entirely in the high frequencies that a clamp flattens first. Once an ordinal loses its ending, the model falls back on the most common way to write that concept, which is a digit.

Where clipping does start to cost you

Two things still go wrong, and neither showed up in this test because the source was clean under the clipping. The first is a clip that was quiet and then cranked up: the noise floor rises with the voice, and the noise is what rewrites sentences. The second is clipping on top of a phone or laptop microphone that already rolled off the low end.

In both cases the damage is not the flat samples. It is the ratio between the voice and everything else that got amplified with it. That ratio is worth checking before you blame the file.

A short burst of clipping is invisible

Clipping in one place behaves differently from clipping everywhere, and not in the direction you would expect. We clipped five half-second stretches spread across the file - the opening word of each sentence - and left the rest untouched. That is 5.28% of all samples flat. The transcript came back 285 characters, 50 words, 50 of 50 matched. Zero edits.

Then we clipped four seconds at the front instead, which is 19.87% of the file but concentrated. That one returned 48 of 50, with both ordinals turned numeric. The same total damage, spread thin, did nothing at all.

• Five half-second bursts, 5.28% clipped: 285 characters, 50 words, 50 of 50 matched.
• Last four seconds clipped, 18.04%: 285 characters, 50 words, 49 of 50 matched, one capital letter.
• First four seconds clipped, 19.87%: 281 characters, 50 words, 48 of 50 matched, both ordinals numeric.

The lesson is that a percentage is the wrong number to look at. What matters is whether the flat part sits on top of a word the model cannot guess from context. A clipped consonant in the middle of a familiar phrase gets rebuilt from the words around it. A clipped address does not, because nothing nearby tells you it was an address.

The other end of the same axis: a file 58 dB under the room

While we had the ladder out we ran the opposite case. We gated the same clip at a threshold above its own speech level, which leaves a file whose loudest sample is 41 out of 32,767. Peak -58 dBFS, average level -74.6 dBFS. You cannot hear it on a laptop at any volume you would tolerate.

It returned the full 285 characters and all 50 words, unchanged. The level had been cut by more than 70 dB and the text did not move.

Push further, to a file of literal zeroes, and the output flips completely. That file returned 1,829 characters of “Thank you very much” and “I'm sorry”, repeated until the model ran out of audio. There is a floor, and it sits somewhere between a signal you cannot hear and no signal at all.

That tells you what clipping and level have in common. Neither one is about how loud the file is. Both are about whether a shape survives at all. A flat shelf at the ceiling and a whisper at -74 dBFS both keep their shape. Zeroes do not.

How to check your own file in two minutes

You do not need a meter to make this call. Upload the file, read the first two sentences, and compare them against what you remember hearing. If the sentences read normally and only names or numbers look odd, the file is fine and you are done.

• Read the text against the audio for the first 15 seconds only.
• If whole clauses are missing, the problem is level or noise, not clipping.
• If only ordinals turned numeric, that is clipping and it is cosmetic.
• If the text repeats one phrase many times, the file has a long silent stretch, which is a different problem.
• Fix the text by hand for the handful of tokens and move on.