Clipped Audio: What It Does to Speech and How Far It Can Be Repaired

What clipping is

Every stage of a recording chain has a maximum level. In a digital file that ceiling is 0 dBFS, the largest value a sample can hold. When the incoming sound would need a larger value, the converter writes the maximum instead, again and again, until the signal drops back. The smooth peaks of the waveform become flat plateaus. That flattening is clipping.

It doesn't only happen in the file. A microphone preamp turned up too far overloads before the signal is ever digitized, producing analog clipping. A plugin chain inside an editor can push the level over the top on export. And a file normalized right up to 0 dBFS and then converted to MP3 or AAC can clip on playback, because lossy encoding changes the waveform slightly and some reconstructed peaks land above the ceiling.

Clipping sounds like crackle, fuzz or a harsh buzz riding on the loudest syllables: laughs, emphatic words, plosives such as p and b, and anyone who raises their voice. Quiet passages in the same file can sound perfectly clean.

How to spot it in a waveform

Open the file in an audio editor and look at the loudest passages.

• Digital clipping shows as peaks that run straight into the top and bottom of the display and stay there, forming flat tops instead of points. Zoom in on one to confirm: a clipped peak is a level line across many samples.
• Many editors can highlight clipped samples, for example through a "show clipping" view option, and some offer a statistics tool that counts samples at maximum level.
• Analog clipping can hide below the ceiling. If the preamp distorted and someone lowered the level afterward, you see squared-off peaks that stop well short of 0 dBFS. Your ears are the better detector here: the distortion is audible even though the meter looks safe.
• In a spectrogram view, clipped moments appear as bright vertical smears reaching high into the frequency range, because clipping adds harmonics that were never in the voice.

Spotting is worth doing before you transcribe anything important. A five-minute audio quality check includes clipping among the problems to look for, along with noise and channel faults.

What clipping does to recognition

Clipping adds distortion at exactly the moments when a speaker is loudest, which are often the moments of emphasis: the conclusion, the punchline, the disputed figure. The added harmonics and buzz cover the detail that separates similar sounds, especially consonants that follow a loud vowel.

The effect scales with severity. A few clipped peaks per minute, typically on plosives or laughs, rarely change a transcript at all. Clipping on every stressed syllable produces scattered substitutions and dropped words. When a speaker is clipped for whole sentences, such as a shouting sports coach or a presenter whose microphone gain was maxed out, recognition can garble or skip entire phrases. Because recognizers tend to produce fluent text, those losses don't always look like errors on the page, a pattern explained in why transcripts miss words.

Why it can't be fully undone

When a sample hits the ceiling, the recording keeps only the fact that the sound was at least that loud. How far above the ceiling it went, and what shape the peak had, are gone. Any repair has to estimate the missing part.

Declipping tools do this by looking at the waveform on either side of each flat top and drawing in a plausible curve. For short clips lasting a handful of samples, the estimate is good, because a short stretch of a waveform is fairly predictable. For long plateaus, the guess becomes unreliable and the repaired peaks can sound smeared or add their own artifacts.

Analog clipping is harder still. The preamp's distortion isn't only flat tops; it bends the waveform as it approaches the limit, and that bending can't be cleanly separated from the voice. Declippers can soften it, but some harshness usually remains.

Repairing clipped speech before transcription

1. Keep the original file untouched and work on a copy.
2. Lower the level of the copy by about 6 dB before declipping. Repaired peaks will be taller than the clipped ones and need room to fit below the ceiling.
3. Apply a declipping tool; free editors such as Audacity include a basic clip-repair effect, and restoration suites offer more sophisticated declippers. Start with conservative settings.
4. Listen to the loudest passages before and after. If the repaired version sounds smeared, metallic or wobbly, reduce the strength or try a different tool.
5. Leave the overall level with peaks around -1 dBFS rather than right at 0 dBFS.
6. Export as WAV or FLAC so a lossy encoder can't push peaks back over the ceiling.
7. Transcribe a minute of the worst passage from both the original and the repaired copy, and keep whichever transcript is more accurate. Repairs that sound better don't always transcribe better.

A worked example: a maxed-out USB microphone

The useful takeaway is where the repair helped and where it didn't. Brief clipping on consonants and laughs was recoverable; sustained clipping on a raised voice wasn't.

What declipping can't do

• Recover a voice that was clipped continuously for whole sentences. The tool is guessing over long stretches and the guess is often audible.
• Remove distortion that happened in an analog preamp or in a phone's built-in processing before the file was written.
• Fix clipping that was later compressed into a low-bitrate file, which mixes clipping distortion with encoding artifacts.
• Separate one clipped speaker from a clean one when both are in the same mixed track. Whatever you apply affects both.
• Replace a retake. If the content can be recorded again, re-recording with correct levels is faster and better than any repair.

Preventing clipping with gain staging

Gain staging means setting the level at each step of the chain so the signal stays well above the noise floor and well below the ceiling.

• Set the preamp or recorder gain while the speaker talks at their loudest, laughing or projecting, not their normal level. Aim for normal speech peaks around -12 dBFS, which leaves room for emphasis.
• Use a limiter on the recorder or interface as a safety net, not as a way to run levels hot.
• Turn on a microphone or interface pad when a loud speaker or a close microphone pushes even minimum gain too high.
• Keep a consistent distance between mouth and microphone; a speaker who leans in to make a point can clip a level that was fine a moment earlier.
• Some field recorders capture 32-bit float files, which makes clipping inside the file very unlikely, though the analog input can still overload with an extremely loud source. Check the recorder's documentation for how it handles input levels.

Recording a little low is the safer error: a quiet recording can be raised later, as covered in transcribing quiet audio, while a clipped one can only be patched. Recorder settings in general are covered in audio recording settings for speech.

Clipped files and mydubly

mydubly transcribes the audio as it arrives and doesn't repair clipping, so any declipping happens in an editor before upload. The browser decodes the audio to 16 kHz mono, splits it into chunks of about 30 seconds and sends compressed chunks over HTTPS for transcription with Whisper. Clipped passages are transcribed as well as the distortion allows, which means the loudest moments are the ones to check.

Upload the repaired WAV or FLAC to audio to text, or a video to video to text. You get a plain transcript, a timestamped transcript and SRT and VTT subtitles, with an optional translated transcript. Use the timestamps to jump to the loud passages you marked during your check. Transcripts cost 1 credit per minute with a 5-credit minimum, so testing a one-minute excerpt of the original and the repaired copy costs 10 credits (1¢) for both.

If you plan a dubbed version, correct the transcript problems you can find first: anything misheard in a clipped sentence is carried into the translation and the synthesized voice.

Distortion often arrives together with noise; speech recognition and background noise explains why aggressive cleanup can backfire, and Whisper hallucinations covers the invented phrases that badly damaged passages can trigger.

Next step: check your loudest minute

Find the loudest minute of your recording, look at it zoomed in, and listen on headphones. If you see flat tops or hear crackle, follow the repair steps on a copy and compare transcripts of both versions. Then adjust your gain for the next recording so the problem doesn't recur. For general recording guidance, see how to improve transcription accuracy.