How to remove background noise from video — locally

The short answer

To remove background noise from a video, run AI denoising on the video file itself, without extracting the audio first. It lowers steady noise such as hum, hiss, air conditioning, and traffic, and a separate dereverb pass reduces room echo. Clipped audio recovers only partly. Locally, use AI Audio Denoise. Picture grain is a separate job: how to denoise video.

The four audio modules

• Audio denoising — steady hum, hiss, air conditioning, traffic bed; keep speech
• Voice enhancement (dereverberation) — untreated rooms, halls, empty apartments
• Vocal separation — isolate dialogue from music, or extract an instrumental
• AI voice change — shift timbre toward a chosen character; preview first

What cleanup can and cannot fix

• Steady broadband noise and room reverb respond well
• Clipped or distorted peaks can be softened, not fully un-clipped
• Dense music mixes separate less cleanly than voice over a bed
• Picture grain is not this tool — use AI Video Quality Repair or the denoise article

How to clean the soundtrack

1. Import the video — or a standalone audio file; no extract first
2. Pick denoise, voice enhancement, or separation; preview to hear it
3. Export locally so the cleaned track stays attached

Match the pass to the noise

Background noise is not one defect. A steady hum, hiss, air conditioner, or traffic bed is a denoise job: lower the constant layer and keep the speech. Room echo from an untreated hall or an empty apartment is dereverberation, which pulls the voice closer without pretending the room was treated. Music under dialogue is separation, and a dense mix separates less cleanly than a voice over a simple bed. AI Audio Denoise runs these as distinct modules on the video file itself.

None of them writes captions. If the next step is SRT or ASS, that is Subtitle Edit, a cue editor. This soundtrack pass does not transcribe speech, translate a line, or burn text into the picture. Picture grain is also a different job. Blocky compression and sensor noise in the image belong in AI Video Quality Repair or the video denoise walkthrough, not in a waveform tool.

You do not extract the audio first. Import the video, or a standalone audio file, and point denoise, voice enhancement, or separation at it. On export the cleaned track stays attached to the picture, so you are not matching a wav back to a timeline by hand. Clipped or distorted peaks can be softened. They cannot be restored as if the microphone had headroom, because clipping threw away the wave at capture. Preview that moment and listen for the limit.

AI voice change is the fourth module, and it is not cleanup. It shifts timbre toward a chosen character. Use it only after the speech is intelligible, and preview before you export, because a timbre shift on a noisy take just moves the noise into a different voice. Order matters: denoise or dereverb first if the room is the problem, then decide whether separation or a timbre change is even wanted.

The free preview plays the opening 5 to 30 seconds of the file and needs no account. If the worst line comes later, a quiet opening can hide a pump or a lisping consonant that only shows up when the person talks over the street, so cut a short test clip that starts on that line. Sign-in and a paid plan are needed to export, convert, repair a file, or download a file from a link. Both run on your machine. The screenshots on this page are real runs on a synthetic voice recording mixed with noise: one denoised, one changed to another voice. Judge your own worst line, not ours.

Listen for what the module is allowed to change. Denoise should leave consonants and breath. If the voice turns watery or the ends of words drop out, ease off and preview a shorter window before you export the whole file. Dereverb should reduce the room, not strip the space until the voice sounds recorded somewhere the room was not. Separation should keep the dialogue intelligible. A leftover cymbal or a missing syllable is a reason to keep the original bed. Run one module, listen, then add another only when the first preview still has a problem that module does not touch. Stacking voice change on a take you have not heard is how a noisy room becomes a noisy character.

• Denoise for steady hum, hiss, and beds. Dereverb for room echo.
• Separation for dialogue against music. Dense mixes stay less clean.
• Voice change shifts timbre. It does not remove noise or write captions.
• Do not extract first. The cleaned soundtrack reattaches on export.
• Listen on the noisy line. Softened clipping is not un-clipping.