What counts as noise, and what does not
A voice note recorded in a car, a lecture taped from the back row, a phone interview with a fan running, a podcast episode where the neighbour’s air conditioner never stopped: these are the recordings people bring to a noise remover, and the first thing to understand is what the tool is listening for.
To a speech-enhancement model, “noise” is sound that is not the voice. Steady sounds, such as fans, engines, air conditioning, traffic hum, a refrigerator, rain and electrical buzz, are the easiest, because they have a texture the model can tell apart from speech in every slice of time. Changing sounds, such as keyboard clatter, dishes, a crowd murmur, doors and footsteps, are harder but still different enough from a voice to be separated most of the time.
Three things are not noise in this sense, and no noise remover handles them well:
- Other voices. Speech is what the model keeps. A second person talking in the background, a television or a crying child are speech-like and stay, sometimes slightly quieter.
- Reverberation. The echo of a voice in a bathroom or an empty hall is the voice itself, bounced. Removing the room means removing part of the speech, and the model treats most of it as voice.
- Clipping and distortion. A recording made too loud, where the waveform was flattened, has lost information; cleaning it removes hiss but cannot restore what was clipped.
Music sits in between: steady background music is partly reduced, but the model is not a vocal isolator and will not strip a song from under a voice.
How the model works and how it was measured
The noise remover on this site uses DeepFilterNet3, an open speech-enhancement model that works on 48 kHz audio. Many times a second, and separately for each band of frequencies, it estimates how much of what it hears is voice, keeps that part and turns the rest down. It also has a second stage that reconstructs the fine structure of the voice in the bands where noise was loudest, which is what keeps cleaned speech sounding natural rather than hollow. The model is about 8 MB, downloads once from this site, and runs on your device; on a recent computer a one-minute recording takes a few seconds to about ten, and phones are slower.
We measured it before shipping it, because “AI noise removal” is a phrase that promises a great deal. The test set was 80 Urdu voice recordings the model had never heard, mixed with 17 kinds of real background noise (fans, traffic, cafés, typing and others) at three levels, from noise as loud as the voice to noise a tenth as loud. The standard measure for this, the signal-to-distortion ratio, went from 4.9 dB in the noisy mixes to 12.7 dB after cleaning: an improvement of 7.8 dB, which in listening terms turns a recording where you strain to follow the words into one where you do not notice the background. For comparison, the classic “spectral subtraction” method used by older audio editors managed 1.35 dB on the same test. Clean speech put through the tool came out almost untouched, so it does no harm to a recording that did not need it.
Those are averages. A steady fan is removed almost completely; a busy café loses most of its murmur but keeps the odd clink; a second speaker stays.
Try it: Remove Background Noise
Drop a recording, listen to Original and Cleaned while it plays, choose Gentle, Strong or Maximum, and download a 48 kHz WAV. The audio stays on your device.
Choosing the strength
The tool offers three strengths, and the right one depends on what the recording is for.
| Strength | What it does | Use it for |
|---|---|---|
| Gentle | Lowers the noise but leaves a little of the room, so the voice keeps its natural texture | Podcasts, interviews, anything people will listen to for pleasure |
| Strong | Removes most of the noise; the voice may sound a little drier | Voice notes, lectures, calls, anything where the words matter more than the sound |
| Maximum | Suppresses everything the model does not take for voice; quiet consonants can soften | Recordings headed for transcription, or so noisy that nothing else helps |
Listen with headphones and switch between the original and the cleaned version while it plays, which the tool lets you do without running the model again. Over-cleaned speech has a recognisable sound: breathy consonants (s, f, th) become dull, and the pauses between words go unnaturally silent. If you hear that, step down a level.
Recording better next time
Noise removal is repair. Ten seconds of thought before pressing record saves more than any model can recover afterwards.
Get the microphone close. Sound falls off quickly with distance; a phone 20 cm from the speaker’s mouth picks up the voice at perhaps ten times the level it does from across a table, while the room noise stays the same. This single change does more than everything else combined.
Turn off what you can. The air conditioner, the fan, the fridge in the next room. Close the window. A steady noise is easy to remove, but a noise that was never recorded is easier still.
Choose the room. Soft surfaces (curtains, a sofa, a bed, clothes in a wardrobe) absorb echo; tiles and bare walls produce it, and echo is the one thing the model cannot remove.
Leave headroom. If the recording app shows a level meter, keep the peaks below the top. A quiet recording can be made louder; a clipped one cannot be repaired.
Record one speaker per microphone when you can. Two people on one phone across a table will both sound distant, and the model cannot separate them afterwards.
Cleaning a recording before transcribing it
The most common reason to clean a recording is to transcribe it, and the two tools on this site are built to be used in that order. Speech recognition models are trained mostly on clean speech, and background noise is the main reason they mistake words; cleaning first, at Strong or Maximum, gives the transcriber a better signal.
The Audio to Text tool then runs OpenAI’s Whisper model in your browser. Three sizes are offered: Fast (68 MB) for clear English and for phones, More accurate (184 MB) for Arabic, strong accents and several speakers, and Best (623 MB, for computers with 8 GB of memory or more), which on our test of 80 Urdu recordings got 85 percent of words right against 74 percent for More accurate. Both tools work on 48 kHz WAV, so the cleaned file goes straight in.
One caution: if the transcript is what matters, do not clean at Maximum and then wonder why short words vanished. Compare the transcripts of the Strong and Maximum versions on a minute of audio before committing to one for a long recording.
Formats, length and privacy
The tool reads the formats your browser can decode: WhatsApp voice notes, MP3, M4A and AAC, WAV, OGG and Opus, FLAC, WebM, AMR and the audio track of MP4, MOV and MKV videos. The cleaning works in mono, as speech models do, and it writes a 48 kHz mono WAV, which is large (about 5.8 MB a minute) but lossless; convert it to MP3 afterwards if size matters. Long recordings are processed in 30-second pieces with a little overlap, so an hour-long lecture is fine on a computer, and limited on a phone mainly by memory.
Nothing you record is uploaded. The model comes to your device; the audio never leaves it, which for an interview, a client call or a medical consultation is the reason to use a tool built this way.
Frequently asked questions
Can it remove other people talking in the background?
Not reliably. Speech is what the model keeps, so a second voice, a television or a crying child stay, sometimes slightly quieter. Music is partly reduced but not removed. Steady and changing non-speech sounds are what it is for.
How much quieter does the noise get?
On 80 recordings mixed with 17 kinds of real noise at three levels, the signal-to-distortion ratio improved by 7.8 dB on average, from 4.9 to 12.7 dB, which turns a recording you strain to follow into one where the background is barely noticed. A steady fan is removed almost completely; a café loses most of its murmur.
Which strength should I use?
Gentle for podcasts and interviews people will listen to, Strong for voice notes, lectures and calls, Maximum for transcription or very noisy recordings. Switch between Original and Cleaned while it plays; if consonants sound dull and pauses go dead silent, step down a level.
Should I clean a recording before transcribing it?
Yes, at Strong or Maximum: speech recognition is trained mostly on clean speech, and background noise is the main cause of wrong words. Compare the transcripts of both strengths on a minute of audio first, because Maximum can soften short words.
What formats does it take, and is the audio uploaded?
WhatsApp voice notes, MP3, M4A, WAV, OGG, FLAC, AMR and the audio of MP4, MOV and MKV videos; it writes a 48 kHz WAV. The 8 MB model downloads once and runs on your device; the recording never leaves it.
Written by Saad Naeem
AI engineer in Doha who builds every tool on ToolboxHere. The figures in the guides come from the official sources named in them or from the tools themselves. Found a mistake? Tell us and it is corrected.