A narration recorded over music, a game cutscene, a clip from a video — and all you want is the voice. Drop the file, press Extract voice, and an AI separation model splits it into the voice and everything else. Listen to both, then save the one you need as MP3 or WAV. The audio is processed on your own machine and never uploaded.
Separating a voice from the music under it used to need a studio plugin or an upload to someone’s server. This page does it with MRX, a separation model from Mitsubishi Electric Research Laboratories trained to split film and video soundtracks into dialogue, music and effects. It runs through ONNX Runtime in a Web Worker on your own processor; the 31 MB model is fetched once, checked against a fixed hash, and kept by the browser for the next visit.
The model never sees the waveform directly. The page turns each 20-second piece of audio into spectrograms at three resolutions — short windows that catch consonants, long ones that resolve the pitch of vowels and notes — and the model returns, for every point of each, the share that belongs to the voice. Applying those shares and transforming back gives the voice; subtracting it from the original gives the background, which is why the two always sum to exactly what you loaded.
Where it stops working is worth knowing before you start. Sung vocals tend to stay with the music, because the model learned singing as part of a soundtrack. Several speakers come out together, since it separates voice from non-voice rather than person from person. And room echo stays with the voice, because it is part of how the voice sounded.
No. The only download is the separation model itself, fetched once from Hugging Face, where it is published, and checked against a fixed SHA-256 hash before it runs. Your audio is decoded by the browser, separated by that model inside a Web Worker on your own processor, and encoded back into a file on your own machine. This site has no server that could receive it. Once the model is cached you can disconnect from the internet and the whole thing still works.
The model, MRX, was trained on thousands of mixtures where the voice, the music and the effects were known separately, and learned what each looks like in a spectrogram — a picture of which frequencies are loud at each moment. For every point in that picture it estimates what fraction belongs to the voice. Multiplying the original by that fraction and turning it back into sound gives the voice track. It looks at the sound at three time–frequency resolutions at once, because a consonant is a short burst and a vowel is a steady tone, and no single resolution sees both clearly.
Because it runs on your computer, on one processor core, instead of on a rented graphics card. Running more cores in a browser requires the site to send isolation headers that GitHub Pages, which hosts this site, cannot send. In exchange your file stays with you, and there is no queue, no account and no daily quota. As a rule of thumb, expect the processing to take about as long as the clip.
Where the music and the voice occupy the same frequencies at the same moment, the information to separate them completely is not in the recording; the model makes its best estimate. In testing, a mix where the voice and the music were about equally loud came out with the voice around eleven decibels above what was left — the background at less than a tenth of its power — which is usually enough that the music is a faint trace. Loud, dense music under a quiet voice leaves more behind than soft background music does.
It measures, for every frequency, how loud that frequency is in the quietest stretches of the voice track — the noise floor that is there when nobody is speaking — and turns each part of the sound down by how close it sits to that floor. Speech, which rises well above the floor, is left alone; steady hiss and hum, which sit at it, are lowered by up to 24 dB. It smooths its decisions across neighbouring frequencies and over time, which is what avoids the warbling artefacts simpler noise gates produce.
The tool places no restriction on its output, and the model is published under the MIT License. Whether you may use the voice is a question about the original recording: separating a voice out of someone else’s video or game does not change who owns that performance. For your own recordings, or material you have the rights to, use the result however you like.