Vocal Remover

It takes a finished song and pulls it back apart, using a separation model that runs on your own device rather than on a server somewhere.

• A karaoke instrumental — the song with the lead vocal taken out, at the original bandwidth and length.
• An isolated vocal — the acapella on its own, for a remix, a practice track or a transcription.
• A four-way split — vocals, drums, bass and everything else as separate files, each from its own model.
• A stem mixer — every stem playing together with mute, solo and level, so you hear the balance before exporting.
• An instant option — a centre-channel method that needs no download at all, for when you want an answer now.
• Nothing uploaded — the audio, the model and the result all stay on your device.

Two stems answers the karaoke question: what did this song sound like before anyone sang on it, and what did the singer sound like on their own. Four stems answers the remix question, and it costs more time and a larger download because each stem comes from its own model.

• Vocals — the lead and anything else the model reads as singing.
• Instrumental — the original with the vocals subtracted, so it keeps every cymbal and every bit of low end.
• Drums — kick, snare, hats and percussion, in a four-way split.
• Bass — the bass line, which a naive vocal remover destroys because it usually sits dead centre.
• Other — guitars, keys, pads, strings - whatever is left once the three above are gone.

On a two-stem split the two files add back up to the original exactly, so you can lay them on top of each other and get the source back sample for sample.

The instrumental is the stem most people came for, and it takes about a minute end to end.

• Drop the song. A video file works too - the audio track is read straight out of it.
• Leave the split on two stems and press Separate. The first run fetches the model once.
• Press Play all and solo the instrumental to hear it on its own.
• If any singing is still audible, the mix was probably wide or double-tracked - try the four-stem split, which models each part separately.
• Save the instrumental, or take both stems as one ZIP.

There are two ways to split a song here, and they fail in different places. The AI path is better at almost everything; Instant exists because a 30 MB download is not always what you want.

• AI stems — a model decides what is voice and what is not, so it works on mono, on wide mixes and on vocals that are not centred.
• Instant — works out what is centred by comparing the two channels, and removes it in proportion. No download, and it finishes in well under a second per minute of audio.

Instant protects the low end and the cymbals on purpose, because bass and kick sit dead centre too - which is exactly what makes older phase-cancellation vocal removers sound hollow.

Every stem is written in the format you choose, at the sample rate the separation ran at.

• MP3 — the one to pick for a karaoke track you are going to play from a phone.
• WAV — 24-bit and lossless, for taking a stem into a DAW.
• FLAC — lossless and roughly half the size of WAV.
• M4A — for Apple devices and anything that prefers AAC.
• OGG — small and open, for the web.

Free covers one song at a time, up to 50 MB, up to six minutes with AI separation or ten minutes on Instant, exported at up to 192 kbps.

Pro adds the four-way split, batches of up to 25 songs, 320 kbps, loudness matching across stems, and lengths bounded by what your device can handle rather than by a number we picked.

• Karaoke — an instrumental of a song that was never released as one.
• Practice — mute your own instrument's stem and play along with the rest of the band.
• Remixing and sampling — an acapella and a drum stem from the same source, lined up perfectly because they came from it.
• Transcription — solo the bass and hear the line without a wall of guitars over it.
• Podcast and interview cleanup — pull a voice out from under background music that was mixed too loud.
• DJ edits — an acapella to lay over a different track.

The song is decoded to raw audio, cut into overlapping windows of about six seconds, and each window is turned into a picture of its frequencies over time. A trained network looks at that picture and writes out the part of it that is voice. Turning that back into sound gives the vocal stem, and subtracting it from the original gives the instrumental - which is why the instrumental keeps the full bandwidth of the source.

All of it happens on your device, using its graphics hardware where that is available and falling back to the processor where it is not. The tool tells you which path it used on each result, because the difference is large and you should not have to guess.

On the reference machine a three-and-a-half minute song separates into two stems in a few seconds once the model is cached, and the instrumental measures 21.6 dB against a known backing track.

• Audio cutter
• Change audio speed
• Pitch shifter
• Audio normaliser
• Audio converter
• Merge audio