How to fix quiet audio — peak normalize in browser, no upload
- Step 1Open the Peak Volume Normalizer — Go to volume-normalizer. The FFmpeg engine loads in the page; no install, no account.
- Step 2Drop your quiet recording — Drag in the faint MP3, WAV, M4A, or voice memo. It is read into the tab locally — a confidential interview never leaves your device.
- Step 3Set how loud you want it — Use Target peak (dBFS). −1 (default) gives a strong, safe listening level; choose −3 if the file is going into further editing.
- Step 4Pick the output format — Keep the original, or pick MP3/WAV/FLAC/M4A. For an interview you will edit later, WAV avoids extra lossy generations.
- Step 5Run the fix — Click run. Pass one measures the quiet peak; pass two applies the boost. The louder file is rendered in seconds.
- Step 6Listen for noise, then download — If the boost revealed hiss, run ai-noise-reducer; otherwise download the
-peaknormfile.
How much boost a quiet file gets
Gain applied = target − measured peak. With the default −1 dBFS target, the quieter the source, the bigger the (honest) boost.
| Measured peak (pass 1) | Target | Gain applied | Outcome |
|---|---|---|---|
| −6 dBFS | −1 dBFS | +5.00 dB | Mild lift — already fairly loud. |
| −12 dBFS | −1 dBFS | +11.00 dB | Clearly louder, still no clipping. |
| −18 dBFS | −1 dBFS | +17.00 dB | Large lift — watch for a rising noise floor. |
| −30 dBFS | −1 dBFS | +29.00 dB | Very faint source; hiss will likely become audible. |
Fixing quiet audio: what to reach for
Peak normalising lifts overall gain. If quietness is uneven or noisy, you need a different sibling tool as well.
| Symptom | Tool | Why |
|---|---|---|
| Whole file is just too soft | This tool (peak normalize) | One measured gain shift brings it up to the ceiling. |
| Boost reveals hiss / hum | ai-noise-reducer | RNNoise speech NN removes the noise the boost exposed. |
| Some words loud, some inaudible | speech-leveler | Compression narrows the gap; gain alone moves both equally. |
| Needs to match a podcast target loudness | loudness-normalizer | LUFS hits a perceptual target, not just a peak. |
Cookbook
Real rescues of faint recordings. Each shows the measured peak from pass one and the exact boost pass two applies — never a blind fixed gain.
A laptop-mic lecture at −19 dBFS
Built-in mics record conservatively. This lecture peaked at −19 dBFS; normalising to −1 reclaims 18 dB of headroom in one shift.
Pass 1 (volumedetect): max_volume: -19.0 dB Target peak: -1 dBFS Gain applied = -1 - (-19.0) = +18.00 dB Pass 2: -af volume=18.00dB Result: lecture now peaks at -1.0 dBFS — clearly audible.
A phone interview, kept lossless
For an interview you will transcribe or edit, export WAV so the boost does not stack a second lossy generation on the original recording.
Input: interview.m4a (max_volume: -14.5 dB) Target: -1 dBFS -> gain +13.50 dB Output format: WAV (pcm_s16le) Output: interview-peaknorm.wav
Leave headroom for editing: target −3 instead of −1
If the rescued file goes into a multitrack edit, a −3 dBFS ceiling leaves room for the mix bus.
Source peak: -16.2 dB Target peak: -3 dBFS Gain applied = -3 - (-16.2) = +13.20 dB Result: peak at -3.0 dBFS, 3 dB headroom retained.
Boost first, then denoise
A very faint field recording at −28 dBFS will surface hiss when lifted. Normalise, listen, then run the denoiser if needed.
Step 1 (this tool): peak -28 dB -> +27 dB -> -1 dBFS (noise floor also rises 27 dB -> hiss now audible) Step 2: /audio-tools/ai-noise-reducer (RNNoise speech NN) Result: loud AND clean.
Why a fixed +10 dB would have clipped
A blind boost is dangerous: a file already peaking at −4 dBFS plus a fixed +10 dB would overshoot 0. The measured approach never does.
Blind +10 dB on a -4 dBFS file: -4 + 10 = +6 dBFS -> CLIPPED This tool (measured to -1): gain = -1 - (-4) = +3.00 dB -> peak at -1.0 dBFS, safe
Edge cases and what actually happens
Recording is near-silent (no real signal)
ExpectedIf volumedetect finds no usable peak, measuredPeak falls back to 0 and the gain is target − 0. There is no signal to amplify; check the mic was actually recording.
Boost raises the noise floor
By designGain is linear: lifting the signal by +20 dB lifts the hiss by +20 dB too. Peak normalising cannot separate signal from noise. Follow with ai-noise-reducer if hiss becomes audible.
Quiet AND uneven (some words inaudible)
By designOne gain shift moves loud and quiet parts together, so the gap stays. To even out level, use speech-leveler (compression), then peak normalise.
Source already clipped despite being quiet on average
PreservedBrief overloads can sit inside a quiet recording. Normalising re-scales; it cannot rebuild clipped peaks. Re-record with more gain if the source is damaged.
Target set to 0 dBFS for maximum loudness
SupportedAllowed (top of the slider) but risky for delivery — inter-sample peaks can exceed 0 dBTP on playback. Prefer −1; for a hard guarantee run true-peak-limiter.
File over 50 MB or 30 min on Free
413 too largeA long lecture may exceed the Free 30-minute cap even under 50 MB. Pro raises it to 120 min / 200 MB; Pro-media and Developer remove the duration limit entirely.
Boosting a lossy file
Re-encodedApplying gain to an MP3/M4A and keeping that format means another lossy generation. Export WAV or FLAC if the recording matters.
Expecting it to make speech intelligible by itself
By designLouder is not clearer. If the recording is muffled or noisy rather than just quiet, normalising will not fix intelligibility — denoise and/or apply voice-eq.
Multiple interviews to fix
Single file onlyThe tool processes one file per run. Fix each recording individually — there is no batch queue for peak normalise.
Re-running the fix
PreservedRe-normalising an already-fixed file re-measures the at-target peak and shifts ~0 dB. It is idempotent, so a second pass does no harm.
Frequently asked questions
How does it make a quiet file louder?
It measures the loudest sample with FFmpeg volumedetect, then shifts the whole file's gain up by exactly enough to put that peak on your target (default −1 dBFS). The boost is calculated, not a fixed guess.
Will it clip my audio?
No, not at or below your target. The math lands the existing peak on the ceiling, so it never pushes a sample past full scale.
How loud can it actually get my recording?
Up to the point where the loudest sample reaches your target. A file peaking at −18 dBFS can rise +17 dB to reach −1. It cannot create level beyond the source's loudest moment.
Why does my fixed file have more background hiss now?
Boosting the signal boosts everything equally, including the noise floor. Run ai-noise-reducer after normalising to remove the exposed hiss.
It is still hard to understand even after boosting — why?
Loudness is not clarity. If the source is muffled or noisy, use voice-eq or ai-noise-reducer; if some words are far quieter than others, use speech-leveler.
Is my interview uploaded anywhere?
No. The boost runs entirely in your browser via FFmpeg WebAssembly. The audio never leaves your device.
What is the best target for a recording I will edit?
Choose −3 dBFS to leave headroom for later mixing. For a finished listenable file, −1 is the default and a good level.
Can I fix several recordings at once?
No — the tool is single-file. Process each recording in its own run.
Which output format should I use?
WAV (pcm_s16le) or FLAC for lossless editing; MP3 or M4A for a small shareable file. Keeping a lossy format means one extra re-encode.
What are the size and length limits?
Free: 50 MB and 30 minutes. Pro: 200 MB / 120 min. Pro-media and Developer: 100 GB and unlimited duration. The duration limit is separate from the size limit.
Does it change the speed or pitch?
No. It only changes gain. For speed or pitch use tempo-changer or pitch-shifter.
Will it overwrite my original?
No. It produces a new file with a -peaknorm suffix; your source is untouched.
Privacy first
Every JAD Audio tool runs entirely in your browser via FFmpeg (WebAssembly) and RNNoise. Your audio files never leave your device — verified by zero outbound network requests during processing.