Pull centred content out of a stereo mix by spectral masking or phase cancellation, with both stems from one pass and a measured report of how far the vocal moved and what it cost everything else.
Initializing in your browser…
Cut a ringtone, notification or alarm from any audio file, snap it to the beat so a repeat does not stutter, and set it to a loudness that holds up against the phone's own tones. Exports .m4r, .mp3 and WAV.
Cut a region out of any audio file, exact to the sample and measured at zero milliseconds of error. Waveform handles, real keyboard shortcuts, click-free cut fades, and WAV or MP3 output that matches its extension.
Convert between MP3, WAV, OGG, AAC, M4A and FLAC with a bitrate, sample rate, channel count and bit depth that all reach the encoder, and a panel that reports what was actually written.
A stereo mix with the lead vocal sitting in the middle needs an instrumental for singing over.
Input
mix.wav - spectral - 40 dB reduction - keep bass below 120 Hz
Output
Vocal 40.02 dB down, panned instruments 0.05 dB, mono sum -3.01 dB
Both stems come out of one pass and the figures above are measured on the exported audio. The build this replaces removed 5.44 dB at its default setting, and 14.14 dB at maximum because its one-pole bass filter still passed 1 kHz 14 dB down; it also cost the hard-panned instruments 3.14 and 5.82 dB, and its output lost a further 40 dB when summed to mono. A vocal delayed between the channels is still beyond any two-channel method, and the tool reports that instead of hiding it.
Pull centred content out of a stereo mix, and see exactly how far it moved and what it cost the rest of the track. Two methods are offered, both stems come out of one pass, and the panel reports what was measured on the audio that came out rather than what the settings asked for. Where the technique cannot work, the page says so with a number.
Take the lead out of a track for singing over, and check on the panel that the backing did not go with it.
Make an instrumental to play or sing along to, keeping the kick and bass with the bass keep control.
The vocals stem is the other half of the same split, so it comes out of the same pass at no extra cost.
The report says how much of the track sits dead centre, which is a useful thing to know about your own mix.
Both methods here separate by comparing the two channels, which is the only handle a browser has without a trained model. A lead vocal is usually mixed dead centre, so it is nearly identical in both, while guitars, keys and effects are spread across the field. Everything below follows from that, including the limits.
The build this replaces was measured before it was touched, on a mix built from a 1 kHz vocal identical in both channels, a 300 Hz instrument hard left, a 2500 Hz instrument hard right, and centred bass. With its default settings it removed 5.44 dB of that vocal, the easiest case this technique can ever be handed, because the Center Width control defaulted to 50 percent and quietly halved the amount subtracted. Pushed to 100 percent it reached 14.14 dB and stopped, because the bass preservation filter was a single one-pole at 200 Hz: at 6 dB per octave it still passes 1 kHz only 14 dB down, and that leakage was exactly the residual vocal. It was not free either. The hard-left instrument lost 3.14 dB and the hard-right one 5.82 dB, and neither is centred at all.
Worse, the output collapsed in mono. Subtracting the centre leaves the sides with opposite signs in the two channels, so summing them cancels the instruments a second time: the hard-panned instrument measured 40 dB quieter in a mono sum. A karaoke track played through a phone speaker was mostly bass. And with the vocal 1.5 dB louder on one side and 0.3 ms late, which is an ordinary thing for a record to be, the setting that removed 14.14 dB removed 2.63 dB.
Spectral is the new default. It transforms both channels, scores every time-frequency cell by how centred it is, and scales that cell. Because both channels are scaled by the same real number, the phase between them is untouched, which is why the result survives a mono sum. On the same fixture it takes the vocal down 40.02 dB while the two panned instruments move 0.05 dB and the bass 0.05 dB, and the mono sum reads -3.01 dB, which is exactly what averaging two uncorrelated signals does and therefore the best any wide stereo can do. With the reduction set to zero the round trip is accurate to -144.5 dB, so the transform itself adds nothing.
Phase cancellation is still offered, because it is instant and on some material it is what you want, but it is fixed. Reduction is now a number of decibels rather than a percentage, and a request for 12 or 24 dB measures 12.0001 and 24.0009 dB on a centred source. The bass split uses a zero-phase Butterworth instead of the one-pole, so nothing leaks back in and asking for full removal takes the vocal below -100 dB. A mono compatible switch outputs the same signal to both channels, which loses the stereo image and stops the mono collapse, and the panel flags the collapse when it is left off.
Both stems, the instrumental and the vocals, come out of a single pass and each gets its own player and its own WAV or MP3 download. The Measured Result panel reports three numbers taken from the output above the bass keep frequency: how far the content both channels carry in phase moved, how far everything else moved, and what a mono sum costs. A mono file is detected on load and the page says plainly that there is nothing to separate, rather than handing back a damaged track. Separation runs in a Web Worker; a six second stereo file takes about 0.2 seconds.
It depends entirely on how the record was mixed, and the tool measures the answer rather than guessing. On a test mix with a dead-centre vocal, the spectral method removed 40.02 dB of it while the hard-panned instruments moved 0.05 dB. With the vocal 1.5 dB louder on one side it removed 21.63 dB. With the vocal delayed 0.3 ms between the channels it removed 0.05 dB, and no setting improves that.
Because both methods separate by comparing the channels, and two copies of a voice a third of a cycle apart are, to that comparison, two different signals rather than one centred one. Nothing about the tuning changes it. A trained model that recognises what a voice sounds like is the way past this, not a better stereo trick.
Phase cancellation subtracts the centre estimate from both channels in the time domain: instant, simple, and it takes a bite out of anything panned, since half of a panned source lands in the centre estimate. Spectral looks at each frequency at each moment, decides how centred it is, and scales only that. It leaves panned material alone and keeps the output mono compatible, at the cost of a little more computation.
Because kick and bass are almost always centred, and removing them leaves an instrumental with no bottom. The Keep bass below control holds everything under it, using a filter steep enough that it does not leak the vocal back in. The one-pole it replaces passed 1 kHz only 14 dB down, which put a hard floor under how much vocal could ever be removed.
What happens when a phone speaker or a club PA adds the two channels together. Content shared by both reads 0 dB, ordinary wide stereo reads about -3 dB because averaging two unrelated signals halves the power, and anything much below that means the channels are cancelling and the track will thin out badly. The classic subtraction reads about -32 dB there unless you switch on mono compatibility.
No, and the page says so as soon as you load one. Both methods work by comparing two channels; with nothing to compare, removing the centre removes the whole track. The warning appears before you process anything.
Reverb, delay tails and doubled takes are deliberately decorrelated between the channels, which is exactly what makes them sound wide, and it is also what makes them survive. What you are hearing is the space around the voice rather than the voice.
Audio is decoded and processed locally with the Web Audio API. Your files are never uploaded to a server.