How we test audio repair
Every check and every fix on this site is proven on real speech that we damaged on purpose, so we know the clean answer. This page shows the recordings, the recipes, the measurements, the pass marks, and the cases the fixes do not solve. Every number here is produced by a test that runs before each change ships.
The short version
Automated tests can prove three things: that a check flags a problem we built into a recording, that it stays quiet on a clean one, and that the fix moves the measurement back toward clean. They cannot prove that every recording in the world will read the same way. Where a check is not sure, it says so in the result instead of guessing, and the fix for that row is left off.
Stereo tests also reverse one channel and check that an audible recording is still measured. Speech levels and spectra average channel power after measuring each side, so opposite polarity cannot erase the evidence. This leaves the recording itself unchanged; it does not correct a cancelling mono mix. See mono versus stereo before changing the channel layout.
Your audio and the limits
Checks and fixes run in your browser in a Web Worker. Nothing is uploaded, which is also why the test suite can run the same code in Node: the Worker and the tests call the same functions on the same decoded samples. The privacy policy lists the little we do measure.
- Fixer pages accept files up to 200 MB and 60 min long.
- The audio editor loads up to five minutes and 50 MB per file and runs a selection repair on up to 30 seconds at a time.
- Decoding uses the browser, so the formats that play in your browser are the ones that open.
Where the voices come from
Two clean readings from LibriVox Short Story Collection Vol. 107, which is in the public domain: 20 seconds from track 3 (Voice A) and 20 seconds from track 4 (Voice B), both 48 kHz, 16-bit, mono. Each sample on the site is a 10 second cut from one of them, degraded by one recipe, and shipped as the file you hear when you press Try a sample. Every damage below is added in software, so the clean original is always available for comparison. None of these is a real bad microphone or a real room, and we say so wherever a page uses them.
How each problem is made
| Sample | Voice | Recipe |
|---|---|---|
| echo.wav | Voice B | Schroeder reverb (four combs, two all-passes) with a 1.6 s RT60 |
| noise.wav | Voice B | Steady pink noise mixed 10 dB under the speech |
| muffled.wav | Voice A | Fourth-order low-pass at 3 kHz |
| distorted.wav | Voice B | Gain +18 dB, then hard-clipped at full scale |
| quiet.wav | Voice A | Gain -22 dB |
| one-ear.wav | Voice B | Stereo file with the right channel all zeros |
| hum.wav | Voice A | 60 Hz plus four harmonics falling at 1/k, 18 dB under the speech |
| wind.wav | Voice B | Gusting brown-noise rumble below 150 Hz, 4 dB under the speech |
| clicks.wav | Voice A | About one click a second: 0.1 to 2 ms bursts of decaying noise, peaking 6 dB under the speech peak |
| plosives.wav | Voice B | Four word starts each get a 120 ms thump at 30 to 60 Hz, peaking 6 dB under the speech peak |
| sibilance.wav | Voice A | High shelf +10 dB above 5 kHz |
| breaths.wav | Voice B | A 0.25 to 0.5 s burst of mouth-shaped hiss at the end of each long pause, 20 dB under the speech |
| dc-offset.wav | Voice A | A constant +2% of full scale added to every sample |
| polarity.wav | Voice A | Stereo file with the voice on both sides and the right side flipped upside down |
| stereo-width.wav | Synthesised music | The sides of a stereo mix turned up 14 dB and the centre down 3 dB, so the sides end up louder |
| speed.wav | Voice A | A music bed 6 dB under the voice, recorded at 44.1 kHz and labelled 48 kHz, so it plays 8.8% fast |
| vinyl.wav | Voice A | Rumble 15 dB under the speech (a wobbling 25 Hz tone plus noise under 35 Hz), 30 tiny ticks a second and white hiss 50 dB under |
| dropouts.wav | Voice A | Four gaps of digital silence, 4, 6, 8 and 10 ms long, each in the middle of a word |
| quiet-hiss.wav | Voice B | White noise 15 dB under the speech, then gain -18 dB |
The recipes live in lib/audio/fixtures/degrade.ts and recipes.ts. A test rebuilds each sample from its recipe and compares it byte for byte with the shipped file, so the file you hear is the file the numbers below were measured on.
What each check measures
Each check turns the recording into one number, then grades it mild, moderate or severe against fixed pass marks. The sentence you see first is written from the grade; the number sits next to it. These are the marks in use today.
| Check | Measure | Mild | Moderate | Severe |
|---|---|---|---|---|
| Echo | s decay, higher is worse | 0.45 | 0.7 | 1.1 |
| Noise | dB SNR, lower is worse | 30 | 20 | 12 |
| Muffled | dB highs, lower is worse | -34 | -40 | -50 |
| Distorted | % clipped, higher is worse | 0.05% | 0.3% | 1.5% |
| Too quiet | LUFS, lower is worse | -27 | -31 | -38 |
| One ear | dB gap, higher is worse | 6 | 10 | 20 |
| Hum | dB hum, higher is worse | -40 | -30 | -20 |
| Wind | % windy, higher is worse | 5% | 15% | 35% |
| Clicks | clicks/min, higher is worse | 8 | 30 | 150 |
| Pops | pops/min, higher is worse | 3 | 10 | 25 |
| Harsh S | dB S, higher is worse | -11 | -7 | -3 |
| Breaths | breaths/min, higher is worse | 5 | 12 | 20 |
| DC offset | % DC, higher is worse | 0.2% | 1% | 3% |
| Out of phase | correlation, lower is worse | -0.3 | -0.6 | -0.85 |
| Stereo width | dB side vs centre, higher is worse | -3 | 0 | 3 |
| Crackle and rumble | ticks/s, higher is worse | 3 | 10 | 30 |
| Dropouts | gaps, higher is worse | 1 | 3 | 10 |
- Noise is speech level minus the level in the pauses, in dB. With no pauses to listen in, the check reports that it could not judge instead of a number.
- Echo is how long the last 10 dB of each fall into a pause takes, in seconds. On synthetic rooms it tracks RT60 within about 15%.
- Muffled is energy from 4 to 10 kHz relative to 300 Hz to 4 kHz, in dB. Clean read speech sits near -24 and some voices read down to about -32, which is why mild starts at -34.
- Distorted is the share of audible samples sitting in flat-topped runs.
- Too quiet is integrated loudness as the LUFS meter reports it.
- One ear, hum and wind are a level gap between channels, hum power relative to loud speech, and the share of the recording with rumble on the microphone.
- Out of phase is how alike the two sides move, from +1 (the same) through 0 (unrelated) to -1 (one side upside down), averaged over half-second blocks with the loud ones counting most.
- Crackle and rumble is small ticks a second, from the same detector as clicks at a lower bar, and the typical level from 5 to 35 Hz against loud sound. Rumble only counts when it stands out from any background hiss and the wind check is quiet.
- Dropouts is the number of gaps where every channel drops to digital silence for 3 ms to a second, with sound right up to both edges. Ordinary pauses keep a faint hiss of room and mic, and fades die away first, so they never count.
What counts as passing
Three rules, and the suite fails if any of them breaks:
- Each sample is flagged for exactly the problems its recipe adds. Flagging anything else on it is a failure, which keeps the checks from crying wolf on each other.
- Both clean readings pass all eight checks untouched, with nothing marked uncertain. Voice A reads 63.5 dB SNR, -23.6 dB highs and -25.1 LUFS; Voice B reads 65.7 dB, -23.6 dB and -24.7 LUFS.
- After the fix chain runs at full amount, every flagged problem reads fine. A fix marked beta only has to improve the grade.
The chain runs in a fixed order, and each step measures the audio it receives from the step before: dropouts, DC offset, wrong speed, out of phase, one ear, stereo width, distorted, clicks, crackle and rumble, pops, hum, wind, noise, echo (beta), muffled, harsh s, breaths, too quiet. Dropouts first, while the gaps are still true silence, then the centre line, then the speed, then channels (a flipped side before anything mixes them) and width, then the waveform (a record's crackle and rumble before the noise fix reads the background), then narrow-band problems, then noise, room and tone, and level last so it reads the finished sound.
Before and after, today
| Sample | Check | Before | After the fix |
|---|---|---|---|
| echo.wav | Echo | severe (1.4 s decay) | fine (n/a) |
| noise.wav | Noise | moderate (15.9 dB SNR) | fine (75.2 dB SNR) |
| muffled.wav | Muffled | severe (-54.7 dB highs) | fine (-20.9 dB highs) |
| distorted.wav | Distorted | severe (10.01 % clipped) | fine (0 % clipped) |
| quiet.wav | Too quiet | severe (-47 LUFS) | fine (-16.5 LUFS) |
| one-ear.wav | One ear | severe (98.7 dB gap) | fine (0 dB gap) |
| hum.wav | Hum | moderate (-21.5 dB hum) | fine (none found) |
| wind.wav | Wind | severe (50.9 % windy) | fine (0 % windy) |
| clicks.wav | Clicks | moderate (60 clicks/min) | fine (0 clicks/min) |
| plosives.wav | Pops | moderate (24 pops/min) | fine (0 pops/min) |
| sibilance.wav | Harsh S | moderate (-4.3 dB S) | fine (-13.5 dB S) |
| breaths.wav | Breaths | moderate (18 breaths/min) | fine (0 breaths/min) |
| dc-offset.wav | DC offset | moderate (2 % DC) | fine (0 % DC) |
| polarity.wav | Out of phase | severe (-1 correlation) | fine (1 correlation) |
| stereo-width.wav | Stereo width | moderate (0.7 dB side vs centre) | fine (-6.3 dB side vs centre) |
| speed.wav | Wrong speed | mild (108.8 % speed) | fine (100 % speed) |
| vinyl.wav | Clicks | severe (504 clicks/min) | fine (0 clicks/min) |
| vinyl.wav | Crackle and rumble | moderate (25 ticks/s) | fine (2 ticks/s) |
| dropouts.wav | Dropouts | moderate (4 gaps) | fine (0 gaps) |
| quiet-hiss.wav | Noise | moderate (17.6 dB SNR) | fine (78.7 dB SNR) |
| quiet-hiss.wav | Too quiet | severe (-42.5 LUFS) | fine (-16.2 LUFS) |
Read the noise rows with care. 75.2 dB SNR after the fix is what steady synthetic noise allows: the noise never changes, so the estimate is nearly perfect. A real room with a fridge cycling, traffic and chair creaks will land far lower. The one-ear, hum and wind rows read none left because the fix removes the thing being measured rather than reducing it.
Measured on . The same numbers are in a CSV file if you want to check them or build on them.
What the fixes do not solve
A page that only shows passes is not a test page, so here is the case we keep in the suite because it fails. Voice A low-passed at 1.5 kHz reads -78.7 dB highs. The muffled fixer brings it to -57.3 dB, still severe. Once the highs are gone, EQ can only lift what is left, and most of what is left above 4 kHz is noise. The 3 kHz sample on this site passes because there is still something there to lift.
- Echo is beta. On the sample it clears, on real rooms with changing distance it can leave a tail, and the tool says beta on the row for that reason.
- Clipping that has been through a lossy encoder loses its flat tops, so the share can read low on a file that still sounds distorted. Clipping vs limiting shows this with waveforms.
- Uncertain findings stay at fine and the fix is not offered: noise with no pauses to listen in, echo masked by a high noise floor, and muffled on files below 16 kHz where there is no band to measure.
- Level fixes read the whole file. Two speakers at different levels are not balanced by it, as this measured example shows.
Meter reference check
We compared 9 generated signals with the same loudness functions the browser Worker uses and FFmpeg's separate EBU R128 scanner. The meter's comparison table includes every result, including differences on abrupt peaks and quiet gaps. This is a small regression set, not a certified compliance test or a guarantee for arbitrary recordings.
To check the independent readings yourself, download the comparison and signal recipes and the Python reproduction script. With Python 3 and FFmpeg installed, run python3 reproduce-meter-reference.py loudness-meter-reference.json in the folder containing both files. It generates 32-bit float WAVs in a temporary directory, measures them locally and prints the comparison. It uploads nothing and plays no audio.
The report records the FFmpeg version, date, sample rates, channels, amplitudes and phase of each signal. FFmpeg's printed readings are rounded to one decimal place. These files let you reproduce that side of the comparison without our application code. The site's fixture tests separately recompute the published fix.fm readings on every change.
Reproduce it
The repository is TypeScript with Vitest. pnpm test runs the whole suite, including the fixture tests that rebuild each sample and recompute every figure on this page from lib/audio/fixtures/how-we-test.test.ts. pnpm samples rewrites the sample files from their recipes. The listening examples in the learn articles are built the same way, each by its own fixture test that also checks the prose quotes the measured numbers.
The audio repair sample library lets you listen to and download the existing before/after files without repository access. It includes cases the repairs leave alone. The numbers on this page come from Node running the same TypeScript the browser Worker runs. The site is built on every change and the suite runs before merging. We check pages by hand in a current Chrome and do not publish a browser compatibility matrix we have not run.
Changes
- October 13, 2026: added the loudness meter's independent reference comparison and public reproduction files.
- October 5, 2026: first version of this page, with the pass marks and results of the current chain.