Split Vocals from Music in Your Browser: AI Source Separation

Two audio jobs that used to require desktop software or cloud services: removing the vocal from a song to make a karaoke track, and cleaning steady background noise off a recording. Both now run in a browser tab, and both work on fundamentally different principles — understanding which one applies to your file is the difference between a clean result and a muddled one.

This guide covers how stereo phase separation pulls centered vocals out of a mix, why it only works on stereo recordings, and how a neural noise suppression network removes hum and hiss without touching the voice.

Two modes, two completely different techniques

The tool offers two operations that are easy to confuse because they both produce “cleaner” audio, but they solve entirely different problems:

• Separate splits a stereo music track into two stems: vocals and instrumental. This is a channel-analysis technique — no AI model is involved, because the information needed to do the split is already in the stereo channels, if you know how to read it.
• Denoise removes steady background noise from a mono or stereo recording. This is a neural-network technique — the model has learned to distinguish speech from noise and applies a gain mask that suppresses the noise while preserving the voice.

The tool auto-detects which mode is appropriate. If you load a mono recording and select Separate, it switches to Denoise automatically — because separation is physically impossible on a mono file, and the alternative would be to produce silence.

Separation: the stereo phase trick

Modern music is mixed in stereo: the vocals, bass, and lead instrument are placed in the centre of the stereo field — equal in the left and right channels. Supporting instruments, reverb tails, and ambient elements are panned to the sides — louder in one channel than the other.

That mixing convention is the key. Anything that is identical in both channels — same amplitude, same phase — is a centre signal. In most modern mixes, that centre signal is dominated by the lead vocal. By analyzing the phase and amplitude differences between the left and right channels, the separation algorithm can isolate the centre and the sides independently:

• Vocals stem — the centre channel, extracted by computing what is common to both channels while cancelling what is different.
• Instrumental stem — the side channels, extracted by removing the centre from the mix. This is the karaoke track: everything except the lead vocal.

The technique is sometimes called centre channel extraction or phase cancellation, and it is the same principle behind the classic “invert one channel and sum to mono” karaoke trick that audio engineers have used for decades. The difference is that a proper implementation does the analysis in the frequency domain rather than just subtracting waveforms, which preserves more of the instruments that happen to share the centre with the vocal.

Why separation only works on stereo

The phase trick requires two channels to compare. A mono recording has one channel — there is nothing to compare against, so there is no way to distinguish the vocal from the instruments by spatial positioning. The tool detects mono input and auto-switches to denoise mode, because trying to separate a mono file would produce silence (the “common signal” of a single channel is the entire signal, so the centre extraction returns everything and the side extraction returns nothing).

This is also why separation quality varies with the source mix. A modern pop track with a clearly centred vocal and wide stereo instruments separates cleanly. An old mono recording remastered into “fake stereo” (where both channels are nearly identical with a slight delay) will not separate well, because there is no real spatial information to exploit. A live recording with instruments bleeding into the vocal’s spatial position will produce a vocal stem with instrumental artifacts.

Denoise: RNNoise and the gain mask

Steady background noise — air conditioner hum, computer fan whir, room tone, tape hiss — is a different problem entirely. It is present in both channels equally and at all times, so stereo analysis cannot separate it from the voice. A neural network can.

RNNoise is a noise suppression model built on a recurrent neural network (a GRU — Gated Recurrent Unit). It was developed by Mozilla’s Jean-Marc Valin for real-time voice communication, and it runs in the browser via WebAssembly. The model works in the frequency domain:

1. The audio is split into short overlapping frames (typically 20 ms each) and converted to the frequency domain via a Fourier transform.
2. For each frame, the neural network examines the spectral content and predicts a gain mask — a set of multipliers, one per frequency band, that say how much of each band to keep and how much to suppress.
3. The gain mask is applied to the frequency-domain signal, and the result is converted back to audio.

The network has been trained on thousands of hours of clean speech mixed with various noise types, so it has learned the spectral signature of human speech. When it sees a frame, it “knows” which frequency bands contain voice and which contain noise, even when they overlap. The gain mask is not a simple on/off switch — it is a continuous multiplier, so a band that is 70% voice and 30% noise gets a gain of roughly 0.7, preserving the voice while attenuating the noise.

This is why RNNoise is called a suppressor rather than a filter. A filter removes specific frequencies permanently; a suppressor adapts frame by frame, removing noise only when speech is not using that frequency band.

What each mode cannot do

• Separation cannot isolate instruments. It separatescentre from sides. If the bass guitar is centred alongside the vocal — which it usually is — the vocal stem will contain bass. True stem separation (drums, bass, vocals, other as four separate stems) requires a spectral or deep-learning model like Spleeter or Demucs, which is a heavier compute load.
• Separation cannot improve a poor mix. If the original track has the vocal panned off-centre, or if the mix is mono, the separation will be imperfect. The result is still usable for rough karaoke or practice, but not studio-quality.
• Denoise cannot remove transient noise. A door slamming, a phone buzzing, someone coughing in the background — these are short, broadband events that look like speech to the model. RNNoise handles steady noise: hum, hiss, fan noise, room tone. Transient interference is a different problem.
• Denoise cannot restore clipped or distorted audio. If the recording is clipped (the waveform hits the ceiling and flattens), the clipping distortion is baked into the signal. The model can suppress the noise around it, but the clipped segments stay distorted.

Recipes

Karaoke track from a stereo song

Load the song, select Separate, and download the Instrumental stem. The vocal stem is also available if you want the a cappella version. Quality is best on modern, well-mixed tracks with a clearly centred vocal — pop, rock, and electronic music from the last two decades typically separate well.

Cleaning a podcast or interview recording

Load the recording, select Denoise. The tool produces a denoised track alongside the original mono for comparison. If the recording has both steady noise and is stereo, you can denoise first, then separate — but for most speech recordings, denoise alone is the right call.

A cappella extraction for sampling or remixing

Separate the track, download the Vocals stem. Be aware that centred bass and kick drum will be present in the vocal stem — this is a limitation of channel-based separation, not a bug. For clean a cappella extraction, a model-based separator is needed.

Privacy, and the thing to check

Audio recordings are often sensitive: unreleased music, confidential interviews, therapy sessions, business calls recorded for note-taking. Cloud audio editors receive the full raw file and process it on their servers. The file, its content, and its metadata exist on someone else’s infrastructure for at least as long as the processing takes — and retention policies are rarely as short as the marketing implies.

Running locally, the audio is decoded by the Web Audio API in the browser’s memory, processed by WASM or JavaScript on your CPU, and the result is encoded to WAV and saved. The only download is the RNNoise WASM module on first denoise use, which is the same for every user and contains none of your data. After that, the tool works offline.

Try it with the Vocal Studio — start with a stereo song you know well, run Separate, and listen to the vocal and instrumental stems back to back. The instrumental stem is the fastest way to hear what the technique can and cannot do: a good separation leaves the song recognizable but distinctly quieter in the vocal range.