Free Vocal Remover (Karaoke Maker)

Worked example

Press "Try a sample". It adds sample-song.wav, about 12 seconds of backing track with a synthesised singer on top, and "Method" starts on "AI separation (best quality)". A panel offers the one-time model download with a "Load model" button, and a line under it says "Expect roughly 0:08 for this 0:12 song on a computer. Once it starts, the bar shows the real time left." (A phone sees 0:57 in a longer sentence.) Press "🎤 Remove vocals". The bar reads "Reading the song…", "Loading the AI model…", "Separating voice and music: part 1 of 2", then "Making the MP3 files…". Two cards appear, "Instrumental (karaoke)" and "Vocals (a cappella)", each with a player, "⬇ Download MP3" and a "WAV" button marked "~2 MB". The files are named sample-song_instrumental.mp3 and sample-song_vocals.mp3. The main button now reads "Re-process".

Which settings to use, and the limits

• Choose AI separation unless you need to avoid the download. "AI separation (best quality)" runs Deezer's Spleeter 2-stems network on your device. It works on most songs, mono ones too, and needs a one-time model download of 18.8 MB (the figure on the "Load model" button also counts engine files that are not cached yet). "Instant (lower quality)" needs no download. It subtracts the centre of a stereo mix between 200 Hz and 6,000 Hz, so it removes what is panned dead centre, and reverb and backing vocals often remain. It refuses mono files. After you change the method, an earlier result is marked as out of date.
• Pick MP3 or WAV for each track. Both tracks are made as 192 kbps MP3 when the run finishes, so "Download MP3" starts at once. The "WAV" button next to it writes a 16-bit stereo 44.1 kHz file when you tap it, and its "~" size is the song length times 44,100 samples times 4 bytes, about 10 MB per minute. Output is always 44.1 kHz stereo, and a mono song is treated as two identical channels. Files are named like song_instrumental.mp3 and song_vocals.wav. In AI mode the vocal track is the song minus the instrumental, so the two add back to the original. If the MP3 encoder cannot load, you get WAV files and a note.
• Drop a video to separate its soundtrack. The drop zone takes audio files and videos such as MP4, MOV, WebM or MKV. For a video the page first shows "Reading the audio from your video…" while the bundled media engine pulls the sound out. Only audio comes back, never a video, and a video with no sound stops with "This video has no audio track to work with." A file above your device's limit is refused and not added: about 2 GB on a normal computer, 1 GB on one reporting 4 GB of memory or less, 400 MB on a phone. A slow-file note starts at 800, 400 and 150 MB.
• Check the time estimate and the length limit. In AI mode a line under the model panel gives the expected time, for example roughly 1:28 for a 4:00 song on a computer or 17:40 on a phone. The phone figure is deliberately pessimistic, and the bar shows the real time left once the run starts. AI mode takes songs up to 20:00 on a computer and 10:00 on a phone or tablet. A long song can add "This song is long for this device's memory and may not finish." On a phone, keep the screen on and the tab open. "Cancel" stops the run, and the next run reads the song again.
• Manage the model download. The model panel says "This tool runs an AI model on your device." Pressing "Load model" or "Remove vocals" downloads it and then separates the song. "Cancel download" stops it with "Download cancelled. Nothing was saved." Once the model is cached the panel reads "Model already downloaded - nothing to fetch." and offers "Remove downloaded model (18.8 MB)". If your device is set to save data, the download is refused until you turn that off or switch to Wi-Fi. On a browser that cannot run the model, the panel says why, and Instant mode still works.

Common problems and how to fix them

It says "The AI model couldn't load. Check your connection and try again, or use Instant mode."

This is the banner for any failure before separation starts: the model download failed, the browser would not store it, or the background worker could not start. The model panel above it stays on screen and gives the specific reason, such as "The download stopped - check your connection and try again." or "There is not enough free storage in your browser for this model.", with a "Try again" button. Fix the cause it names and press "Try again", or switch to Instant, which needs no download.

It says "The AI separation stopped before it finished, which usually means this device ran out of memory. Try a shorter song, close other tabs, or use Instant mode."

The separation worker reported an error while working, or sent no message for 3 minutes, which the page treats as stuck (an engine download in progress counts as activity). The model needs about 400 MB of memory on top of the song itself, so a long song on a phone is the usual cause. Use a shorter part of the song, close other tabs, or choose Instant. The failed worker is thrown away, so pressing "Remove vocals" again starts clean.

It says "This file couldn't be read as audio. Try an MP3, WAV, M4A or FLAC file." or, for a video, "Could not read the audio from this video. Try an MP4 or MOV file, or save the audio as MP3 first."

The browser's own audio decoder could not open the file, so there was nothing to separate. Save the song as MP3, WAV, M4A or FLAC and add it again. For a video, the media engine failed to pull the sound out: re-export it as MP4 or MOV, or save the audio as MP3 first. A file with under one second of audio gives "This file is too short to separate. Use a clip of at least one second." instead.

It says "This song is longer than 20:00, the most this device can separate in one go. Trim it into shorter parts first." or "This file is 2.6 GB, which is over the ~2 GB this device can process in the browser. Try a smaller file or trim a shorter clip first."

AI mode holds the song's audio in memory while the model works, so it takes up to 20:00 on a computer and 10:00 on a phone or tablet; the time in the message is your device's own cap. The check uses the length the player reports and applies to AI mode only. Cut the song into parts and separate each. The second message is the file-size limit (2 GB, 1 GB or 400 MB by device), and that file is not added.

It says "This file is mono (or both channels are the same), so Instant mode has nothing to cancel. Use AI separation, which works on mono songs."

Instant mode works by telling the centre of the mix from the sides. In a mono file the two channels are the same, so there is no difference to separate. A stereo file counts as mono when the side signal is 40 dB below the middle or lower. Switch "Method" to "AI separation (best quality)". A related warning, "This song has little centre-panned content: the result is about 34 dB quieter than the original...", appears when an Instant result is over 30 dB quieter than the input.

How it compares

The AI mode makes two files, an instrumental and a vocal track, with Deezer's Spleeter 2-stems model (MIT licence). The page itself warns to "expect a faint trace of voice in loud passages", and instruments that sound like a voice can end up in the vocal track. In this site's own tests it separated clearly better than Instant mode on every test song. It splits into two tracks only; for drums and bass as separate tracks, use Stem Splitter (Vocals, Drums, Bass). Because it runs in the browser, it needs the model download and takes roughly a third of the song's length on a computer, far longer on a phone, with a 20:00 cap per song. Nothing is uploaded.