Free tool · AI noise model · Runs in your browser
Remove background noise from video
Short answer
The tool cleans the soundtrack with the RNNoise speech model and copies the picture exactly as it is. On a 10 second 1080p clip with street noise on its soundtrack, the video stream was bit for bit identical afterwards (same checksum), the audio SDR rose by +3.9 dB, and the file went from 13,533,148 to 13,598,100 bytes (+0.48%). Nothing is uploaded.
Nothing is uploaded: processing runs on your device. Maximum file size 100 MB. The first run loads a ~30 MB engine and a 298 KB noise model, once per visit. The picture is copied unchanged; only the sound is re-encoded.
Key takeaways
- The picture is not re-encoded: the video stream's checksum was the same before and after (identical for both settings), so there is no generation loss and no wait for a video encode.
- Only the sound is re-encoded, as AAC at 192 kbps in an MP4 file with the index at the front for streaming. File size moved by +0.48% with Gentle and +0.38% with Standard.
- On the test clip the audio SDR rose by +3.9 dB with Gentle and +4.0 dB with Standard, starting from 6.4 dB. The audio page has the full results for two noises at four levels.
- Processing the 10 second clip took 0.6 seconds in a headless run of the same engine, because only the audio is processed.
- The video needs an audio track. If the picture codec cannot be stored in MP4, or the file has no sound, the tool reports a failure instead of writing a file.
What changes in the file, and what does not
The tool keeps the first video stream and the first audio stream. The picture goes through unchanged and the audio goes through the noise filter and a new encode.
| Before | After | |
|---|---|---|
| Video stream | H.264, 1920 × 1080, MD5 734daafe6530… | Copied, MD5 734daafe6530… (identical) |
| Audio stream | AAC at 128 kbps, speech with street noise | Cleaned, AAC at 192 kbps |
| Container | MP4 | MP4, index at the front |
| File size | 13,533,148 bytes | 13,598,100 bytes with Gentle (+0.48%), 13,584,307 with Standard (+0.38%) |
Audio at 192 kbps is slightly larger than the 128 kbps of the test file, which is why the file grows by half a percent. A video with a larger soundtrack will move differently.
Measured on a real clip
The test clip is a 10 second 1080p excerpt of Big Buck Bunny (CC BY 3.0, Blender Foundation) with the speech and street-noise mix from our audio test as its soundtrack, at 5 dB noise. The first 10 seconds of that mix are a little cleaner than its 30 second average, which is why the starting SDR is 6.4 dB rather than 5 dB. The excerpt is animated, not camera footage, but the picture is only copied, so that does not affect these results.
| Setting | Audio SDR before | Audio SDR after | Change | File size | Time |
|---|---|---|---|---|---|
| Gentle | 6.4 dB | 10.3 dB | +3.9 dB | 13,598,100 bytes | 0.6 s |
| Standard | 6.4 dB | 10.4 dB | +4.0 dB | 13,584,307 bytes | 0.6 s |
The two settings scored almost the same on this clip, as they did at 5 dB on the audio page (+4.7 to +4.8 dB and +4.9 to +5.1 dB there).
Why the picture stays untouched
Stream copy hands the video packets to the new file without decoding them. That is why the checksum of the video stream is the same afterwards, and why the job is fast: the cost is only the audio, which took 0.6 seconds for the 10 second clip.
It also sets the one hard rule. The video stream has to be storable in MP4. H.264 and HEVC from phones and cameras are; other codecs may not be, and then the tool stops with an error instead of writing a damaged file.
When this is the right tool
- A phone clip with a fan, air conditioning or traffic behind a talking person.
- An interview or screen recording with a computer fan under the voice.
- You want the cleaned sound without re-encoding the picture, so there is no quality loss in the video.
When it is not
- Music videos and performances: the model is built for speech, and we did not test music.
- Echo in a large room: the model's notes list fans, air conditioning and computers as the intended noise.
- A video with no audio track, or with the sound in a second audio stream: the tool uses the first one.
Free tools:Remove background noise from audioRemove the sound from a videoExtract MP3 from a video
Questions people ask
Does removing noise from a video lower the picture quality?
No. The picture is copied, not re-encoded. On our test clip the video stream had the same checksum before and after.
Does the video get uploaded?
No. The tool runs in your browser. Files up to 100 MB are accepted, and the first run downloads an engine of about 30 MB and a model of 298 KB, once per visit.
Why did my file get slightly larger?
The soundtrack is re-encoded as AAC at 192 kbps. On our clip the original was 128 kbps, so the file grew by +0.48%.
Which setting should I choose for a video?
Gentle for most clips. It keeps the voice natural. Use Standard when the noise is nearly as loud as the voice.
What if my video has no sound?
The tool needs an audio track and reports a failure without one.
Can it fix echo or music in the background?
It was built for speech with steady noise such as fans and air conditioning. We did not test music or echo.
What format do I get back?
An MP4 with the original picture and a cleaned AAC soundtrack.
Related
Sources and how we checked
- FFmpeg Filters Documentation: arnndn, afftdn and asdr (read 2026-10-06)
- GregorR/rnnoise-models: README and model notes (somnolent-hogwash) (read 2026-10-09)
- Internet Archive: The Heart of a Mystery, LibriVox recording (public domain mark) (read 2026-10-09)
- Wikimedia Commons: 20220623 Noise pollution, residential area, by WrS.tm.pl (public domain) (read 2026-10-09)
Reviewed October 9, 2026. The measurements ran the tool’s own filter arguments on the same single-thread WebAssembly engine your browser downloads. The scripts, the results and the source recordings’ licenses are in data/seo/hub/evidence/2026-10-09/noise.