Audio Enhancer

Clean up a voice recording: take out steady background noise such as a fan, traffic, hum or room hiss, even out the volume, listen to the original and the result side by side, and save it as MP3 or WAV. The file stays in your browser.

Drop your recording here or click to browse

MP3, M4A, WAV, AAC, FLAC, OGG, OPUS, WEBM or MP4, up to 100 MB and 30 minutes.

The file is processed in your browser and is not uploaded.

How to remove background noise from audio

Three steps, in your browser.

  1. 1

    Drop the recording

    MP3, M4A from a phone, WAV, FLAC and other common formats, up to 30 minutes. The page reads it and shows its length and how loud the quiet parts are: the background level.

  2. 2

    Enhance, then set the strength

    Press Enhance and the noise model runs through the recording, a few seconds per minute of audio. Then switch between Original and Enhanced while it plays and move the strength slider until the voice sounds natural.

  3. 3

    Download

    Save the result as an MP3 or a 16-bit WAV. Normalize is on by default, so the cleaned file also comes out at a sensible level; turn it off to keep the original volume.

Only too quiet, not noisy? That is a different job: use the Volume Booster.

What the audio enhancer does, and what it does not

  • Steady noise under speech

    The model was trained to separate a voice from the noise around it: air conditioning, a fan, a running engine, traffic, electrical hum, the hiss of a cheap microphone. That is where it does its best work.

  • Strength is a mix, not a switch

    At 100% you hear the model's full output. At 60% you hear 60% cleaned and 40% original, which keeps a little room tone and often sounds more natural on a quiet recording. Change it as often as you like; nothing is recomputed.

  • Normalize, if you want it

    After the noise is gone the recording often sounds quieter than it is. Normalize lifts it so the loudest peak sits at -1 dB, the same rule the Volume Booster uses. It is one fixed gain: nothing is compressed.

  • What it cannot fix

    Echo from a bare room, two people talking over each other, a voice that clipped when it was recorded, wind hitting the microphone, music you want to keep. Those are not steady noise, and the model will either leave them or damage the voice trying.

Recordings that usually need it

  • A site visit or plant tour

    Machines, ventilation and traffic sit under everything the customer said. Run the enhancer at full strength first; if the voice sounds thin, back off to 70 or 80%.

  • Notes dictated in the car

    Road and engine noise is as steady as it gets, so the model handles it well. Normalize afterwards: dictation in a car is usually quiet.

  • An interview in a café

    The espresso machine and the ventilation come out; the people at the next table mostly do not, because they are speech too. Expect an improvement, not silence.

  • A meeting recorded with the phone on the table

    Projector fans and air conditioning go, and the far voices get clearer once the hiss under them is gone. If the far voices are still too quiet, run the result through the Volume Booster.

How the noise removal works

The page uses RNNoise, an open-source noise suppressor from Mozilla and Xiph.Org. It is a small neural network trained on clean speech mixed with separately recorded noise. Working on 10-millisecond slices at 48 kHz, it estimates for each of 22 frequency bands how much of the sound is voice and how much is noise, and turns the noise bands down while keeping the voice bands as they are.

Everything runs in your browser through WebAssembly: the model is downloaded once (about 2 MB) and your recording never leaves your computer or phone. A 10-minute file takes well under a minute on a recent laptop; the progress bar shows where it is.

The page also measures the background level before and after: the average level of the quietest tenth of the recording. A change of 10 dB or more is clearly audible; a change of 2 dB means the noise was not steady enough for the model, or there was little noise to begin with.

Cleaner is easier to hear. Text is easier to use.

You cleaned one recording by hand so you could hear it. ParrotNotes records the call, visit or interview on your phone and transcribes it with Deepgram speech-to-text, which copes with a moderately noisy room, so you get the transcript, a summary and the action items without a cleaning step. Free for 100 minutes of recording a month.

Get ParrotNotes free

Related tools

Audio Enhancer FAQ

Sources