Clipped audio:
Can you save the take?
Lightly clipped audio may be worth trying to repair on a copy; if the damage remains unacceptable and you can record again, rerecord. A correct MIDI note does not prove the waveform is restored. Our unclipped and hard-clipped tones produced identical MIDI files despite different waveforms. Decide whether you need usable recorded sound or only a checked note draft—not just whether conversion succeeds.
Download the clipping decision table (CSV)Different waveforms. The same note file.
On October 9, 2026, at 11:54 Asia/Seoul, this Windows machine, ATMN-WIN-15fa5aa9, generated a known-pitch A4 sine tone and a deliberately hard-clipped version of the same sample sequence. We passed both through this site's unchanged production analysis functions in Balanced mode. These are synthetic, code-level measurements, not a recorded performance or browser upload test.
| Input | WAV length | Detected pitch | Note events | Detected event length | Rail samples |
|---|---|---|---|---|---|
| Unclipped tone | 2.000 s | 69 · A4 | 1 | 1.811 s | 0 |
| Hard-clipped tone | 2.000 s | 69 · A4 | 1 | 1.811 s | 67,400 |
Swipe the measurements sideways. With a keyboard, focus the table and use the arrow keys.
“Rail samples” counts actual signed PCM values equal to the positive or negative limit. It is not a listening score or a universal clipping threshold. Here the clipping is known because we injected it. Event lengths above are rounded detector values before MIDI serialization; the exact value for both is 1.811156462585034 seconds, not the full WAV length.
Both unedited MIDI outputs are the same 42 bytes, byte for byte. Their SHA-256 is 445e24c302406b802b2fa81c9f16c35afcce405f7d497ed55a68c2f1fb87f30c. That proves equality of these two outputs, not recovery of the original audio or accurate performance on clipped music.
- Unclipped waveform fingerprint · exact PCM SHA-256
d81686d3fa0242b7e9e325296956d45f1e84918b69f2e5864efca01f904696c5- Hard-clipped waveform fingerprint · exact PCM SHA-256
7a6dcdd5f27e1789e0a38105b4742cf1c5784cad2f5944d07d888cf9b772987e
The hard-clipped file contains a longest uninterrupted run of 39 rail samples; its sample range is −32,768 to +32,767. The clean range is −11,468 to +11,468. A fresh repeat reproduced both WAV and both MIDI hashes without changing the production source. Fingerprints verify file identity, not how a recording sounds.
Rerecord, keep a note draft, or try repair?
Download the decision table as CSV to keep the observed cases and recommended next actions together. The first cases are measurements; the other rows are guidance, explicitly marked as unmeasured. A recommendation is not a repair result.
| Input condition | Observed result | Next decision |
|---|---|---|
| Our unclipped tone Measured October 9 | No rail samples; one A4 event. | No clipping repair indicated for this fixture. Its shortened event still needs checking. |
| Our hard-clipped tone Measured October 9 | 67,400 rail samples; the same A4 event and MIDI bytes. | Do not approve the recording from MIDI alone. Use the clean original, or rerecord if damaged sound is unacceptable and recording again is possible. |
| Your only take, possibly lightly clipped Recommendation, not tested here | Not measured. | Try a suitable repair on a copy and compare the sound. Accept only if the result meets your actual use; otherwise seek a clean source. |
| You need notes, not this recorded timbre Recommendation, not tested here | Not measured. | Try a short transcription and verify pitch and timing against a reliable reference. A usable draft does not make the recording repaired. |
- Choose the deliverable. If the recording itself will be heard, inspect its sound separately from its MIDI. If you only need the melody, check the resulting note draft; do not use file creation as the acceptance test.
- Keep an untouched reference. Use a clean original or another take when available. Turning the damaged file down cannot demonstrate that missing peaks have been reconstructed. This run did not test a volume-only fix.
- Test repair on a duplicate, not the only take. Audacity's Clip Fix documentation describes interpolation and limits its likely usefulness to lightly clipped material. That is a documented option, not a repair we ran or a guarantee for the severe fixture above.
- Make a separate audio and note decision. Compare the repaired copy with the original or a clean reference, then inspect any new MIDI. If neither result meets the intended use, choose a clean source or rerecord. We have not measured a repair success rate or a universal severity cutoff.
Do not use “67,400” as a pass/fail limit for your recording. It is the count in this deliberately constructed file. An unknown file's peaks, normalization or earlier processing can differ; absence of samples exactly on today's rails does not certify that it was never clipped.
Download the comparison and its record
- Rerecord-versus-repair decision table (CSV).
- Unclipped WAV input · Unedited MIDI output.
- Hard-clipped WAV input · Unedited MIDI output.
- Exact events, dates, machine identifier and complete hashes · Reproduction script.
Method and boundaries
The original measurement ran October 9, 2026, from 02:54:30.339Z to 02:54:31.054Z on ATMN-WIN-15fa5aa9, Windows x64, Node v24.19.0. The repeat ran from 02:54:31.849Z to 02:54:33.411Z. These timestamps are UTC; the table's calendar date uses Asia/Seoul. The identifier distinguishes this machine without publishing its hostname.
Each WAV contains 88,200 mono frames of signed 16-bit PCM at 44,100 Hz and is 176,444 bytes. The generated source is 440 Hz, amplitude 0.35 before integer rounding, with 8 ms edges. We multiplied the exact clean PCM integers by 8 and clamped them to −32,768 and +32,767 to create the second file. This is a gain-and-clipping perturbation, not a loudness-matched isolation of clipping alone.
We parsed the actual WAV integers into Float32 samples, called the unchanged mixToMono, downsample, detectNotes and buildMidi functions from the production source, and saved the unedited outputs. Both analysis arrays contain 22,050 samples at 11,025 Hz. The JSON records the full source-code hash and exact events; the detector was not tuned to this result.
In the site project, run node tests/clipped-audio-research.mjs --check to repeat without overwriting the original public observations. The downloadable reproduction script is the same code and expects this project's dist/script.js; it is not a standalone de-clipping tool.
This comparison bypasses browser decoding and the file picker. MDN documents AudioContext-rate resampling during Web Audio decoding; an end-to-end browser result need not match this direct-parser experiment. We did not measure listening quality, repair output, a real voice or instrument, a full song, a DAW import or other converters. The material did not come from a video.
The new observation here is narrow: our injected hard clipping changed the PCM fingerprint and rail count but left this isolated tone's MIDI bytes unchanged. It does not establish that clipped recordings generally transcribe correctly, that a de-clipper restored peaks, or that every wrong note comes from clipping.
If the problem instead shifts the entire source's pitch and speed, use the separate sample-rate-label check. If the note result is empty, investigate channel cancellation or insufficient analysis windows. Those are different mechanisms, not interchangeable clipping diagnoses.