Remove voice from video
Upload a video or audio file, and get it back without the voice, with the music and background sound kept.
Remove the voice from your own video
Sign in free to remove the voice from videos and audio up to 3 minutes. Paid plans take files up to 10 minutes.
Example: the same clip with and without the voice
An 18-second clip with a man reading over music, made from public-domain and openly licensed recordings. The second video is the tool's own output, unedited. We mixed the recordings ourselves, so we could measure the result: the voice left in it is about 60 dB quieter than in the original, and the music stayed within 1 dB of its level.
Original, with the voice
Voice removed, music kept
Narration: the opening of Moby-Dick by Herman Melville, read for LibriVox (public domain). Music: "Inspired" by Kevin MacLeod (incompetech.com), licensed under Creative Commons: By Attribution 4.0. Footage: views of Earth from the International Space Station, NASA (public domain).
How to remove the voice from a video
- Sign in free and open Remove voice. Speech removal needs an account because it runs on a small shared pool of servers.
- Choose the video or audio file. Free accounts take files up to 3 minutes, paid plans up to 10 minutes.
- Choose Remove the voice. The model takes about 20 seconds per minute of audio, plus up to a minute when the GPU server has to start.
- Watch the result on the page, then download the MP4 without the voice, the music and background sound as WAV, or both.
How the voice is removed
The tool uses BS-RoFormer, a vocal separation model trained by viperx and shared in the public model set of Ultimate Vocal Remover (UVR). It runs on a GPU with the MIT-licensed audio-separator library. The model splits the sound into the voice and everything else, and the tool keeps everything else. For a video, that sound replaces the original track, and the picture is copied as it is when MP4 can hold it.
We picked it after testing seven free separation models on a real voice over real music. BS-RoFormer left the least voice behind: about 20 dB less than the model we used before. If the GPU server is unavailable, the tool falls back to that smaller model, and more of the voice can remain.
It works best on clear speech over music. Crowd noise, choirs and speech over loud effects can still leave faint traces. Listen to the whole result before you use it.
When you do not need a voice remover
- If the voice-over is on its own track in your editor project, mute or delete that track. Nothing has to be separated.
- If you recorded with two microphones on two channels, split the channels and keep the one without the voice. The stereo splitter does this in your browser.
- Adobe Premiere's Enhance Audio separates dialogue, music, ambience and sound effects into layers with their own volume, according to Adobe's help page.
Replace the voice instead
To put a new voice on the video, the voice-over tool reads a timed script over the sound you keep. To change the language, video dubbing removes the original voice the same way, translates the speech, and lays a new voice over the music.
Limits
- No account: watch the example on this page. Removing the voice from your own file needs a free account, because speech separation runs on a small shared pool.
- Free account: files up to 3 minutes and 100 MB, within 90 minutes and 50 processing jobs a day.
- Paid plans: files up to 10 minutes, with more processing a day (Starter 6 hours, Pro 30 hours).
- A video counts its length twice and uses two jobs (removing the voice, then putting the sound back on the picture). An audio file counts once and uses one job. Limits reset at midnight UTC.
Acceptable use
Remove voices only from files you made or have the rights to edit. The music and sound left in the file keep their own copyright: removing the voice does not give you the right to reuse someone else's soundtrack.
Related tools
FAQ
Is it free? +
Yes, with a free account, for videos and audio up to 3 minutes. Paid plans take files up to 10 minutes and include more processing a day. Without an account you can watch the example on this page; removing the voice from your own file needs an account because speech separation runs on a small shared pool of servers.
Can I remove just one person's voice? +
No. The model targets all voices together, so it cannot keep one speaker and remove another, and some vocals can remain. If each person was recorded on their own track or channel, mute that track in your editor instead.
Does it remove singing too? +
Yes, usually. BS-RoFormer is a vocal model, so it removes sung vocals along with speech and keeps the instruments. Faint traces can remain in hard cases, such as a choir or vocals with heavy effects.
How long does it take? +
About 20 seconds per minute of audio, plus up to a minute when the GPU server has to start. Putting the sound back on a video usually takes seconds, and longer when the picture has to be converted to H.264, for example from WebM.
Which files work? +
MP4, MOV, WebM, MKV and most other video files with sound, and MP3, WAV, M4A and other audio files, up to 100 MB.
What do I get? +
For a video, an MP4 with the same picture and the music and background sound without the voice. For video and audio, the separated music and background sound as a WAV file you can use in an editor.
Where is my file processed? +
The file is uploaded to SoundHalo on Cloudflare. For speech separation, only the audio track is sent to a GPU server SoundHalo rents from Modal, and it is deleted there when the request finishes. Mixing runs on SoundHalo's own servers. Access to the files expires after 30 days (longer on paid plans, up to a year), and you can delete them from the page.
What are the limits? +
No account: watch the example on this page. Removing the voice from your own file needs a free account, because speech separation runs on a small shared pool. Free account: files up to 3 minutes and 100 MB, within 90 minutes and 50 processing jobs a day. Paid plans: files up to 10 minutes, with more processing a day (Starter 6 hours, Pro 30 hours). A video counts its length twice and uses two jobs (removing the voice, then putting the sound back on the picture). An audio file counts once and uses one job. Limits reset at midnight UTC.