Loading tool…
Background noise turns a good recording into a tiring one. Drop in a voice recording or a video and RNNoise, a small neural noise suppressor, cleans the speech in seconds on your own device. You can compare before and after in sync, then download clean audio or the same video with a clean soundtrack.
Processing locally on your device
Overview
Classic noise reduction learns a noise profile from a quiet passage and subtracts that spectrum everywhere. That handles steady hiss, but when pushed it sounds watery or metallic, and it can't follow noise that changes, such as traffic, a passing car or a dishwasher starting up. A neural suppressor works the other way around. It has learned what speech sounds like, and for each 10-millisecond frame it estimates how much of each frequency band is voice and how much is noise.
This tool uses RNNoise, created by Jean-Marc Valin at Xiph.Org (the Opus codec team) and released under the BSD licence. It mixes classic signal processing with a small recurrent neural network. Your file is resampled to 48 kHz, cut into frames of 480 samples and cleaned channel by channel. The engine is built into the page as WebAssembly, so there is no model download and no graphics card is needed. In our tests it processed audio about six to eight times faster than real time per channel, so a 10-minute mono interview takes well under two minutes. RNNoise adds 20 ms of delay, which the tool removes, so the cleaned track lines up with your video and with the original.
Three controls let you shape the result. Strength blends the cleaned signal with the original: 100% removes the most, and 60–80% keeps some natural room tone so the voice doesn't sound processed. Silence the gaps uses RNNoise's built-in voice-activity detector to turn down everything between phrases, which clears breaths, keyboard taps and low hum. You set the sensitivity and how far the gaps drop, and the gated parts are shaded on the waveform. A Before/After switch flips between the original and the clean version while playback continues, so you hear exactly what changed.
For video files the soundtrack is cleaned, and the tool can put it back into the video with FFmpeg without re-encoding the picture, so quality is unchanged and it's quick. There is also a live microphone mode. An AudioWorklet sends your microphone through RNNoise with roughly 50 ms of delay, with level meters, an on/off switch for A/B listening and optional monitoring on headphones. Everything you say is recorded clean and can be downloaded when you stop.
| Noise type | Result | What helps |
|---|---|---|
| Computer fan, AC, room tone | Excellent | Default settings |
| Electrical hum and hiss | Good | Full strength, gap gate on |
| Traffic and wind rumble | Good | Full strength |
| Keyboard and mouse clicks | Fair to good | Gap gate for pauses between phrases |
| Background music or TV | Limited | Vocal remover instead |
| Other people talking | Poor | Re-record or edit those parts out |
Drop in a voice recording (MP3, WAV, M4A, OGG, FLAC) or a video (MP4, MOV, WebM, MKV).
The audio is cleaned automatically; play it and switch between Before and After.
Adjust Strength, and turn on Silence the gaps if you want quieter pauses, then click Apply changes.
For a video, keep Keep the video on to get the same video back with clean sound.
Choose an audio format if needed and download the result.
Or switch to Live microphone to clean and record your voice in real time.
Use cases
Clear the hum of a computer fan or air conditioner from a home-studio recording before editing, so the episode sounds like it was made in a quieter room.
Drop the whole MP4 in, clean the laptop-microphone noise and download the same video with better sound, without waiting for the picture to be re-encoded.
Reduce traffic, wind rumble and café noise in recordings made on a phone, where doing a second take isn't possible, and make transcriptions more accurate.
Use the live mode to hear how much noise your room and microphone pick up, compare suppression on and off on headphones, and record a clean test take.
Pro tips
Answers to common questions about the ai noise suppression.
Steady or slowly changing noise: computer fans, air conditioning, hum, hiss, road rumble and room tone. It also helps with keyboard clicks. It is weaker against music, TV and other people talking, because those share characteristics with the speech it is trained to keep.
Not well. RNNoise is trained to keep speech and suppress everything else, so it can treat instruments as noise. For music, use a spectral noise reducer, or the vocal remover if you want to separate a voice from a backing track.
Yes. RNNoise delays its output by 20 ms, and the tool compensates for that exactly. When you keep the video, the picture stream is copied untouched and only the soundtrack is replaced.
RNNoise estimates the probability that each 10 ms frame contains speech. With the gate on, frames without speech are turned down by the gap level you choose (for example −40 dB). The gate opens slightly before speech starts and stays open for 150 ms after it ends, so words aren't chopped.
An AudioWorklet passes your microphone to RNNoise in 10 ms frames and plays the cleaned frames back with about 50 ms of delay. The browser's own noise suppression is switched off so the comparison is fair. The clean signal is kept in memory and offered as a download when you stop.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The ai noise suppression is free to use with no sign-up, no watermarks and no daily limits.