You want to make a karaoke version of your favorite song, or you need the isolated vocals (acapella) for a remix, or you just want to hear only the drums of a track to study a drummer's pattern. All three cases require the same thing: splitting a mixed song into its individual components. Until a few years ago this was practically impossible without access to the original studio tracks — today an AI model does it in the browser, without uploading the song anywhere.
Why splitting a mix is a hard problem
When a song gets mixed, the vocals, drums, bass and remaining instruments get combined into a single audio wave — mathematically, their signals get summed into the same channel. Undoing that sum to recover each instrument separately is an ill-posed problem: there are infinitely many combinations of "vocals + drums + bass + other" that would add up to the same final mix. There's no direct mathematical formula that solves it; it takes a model trained on huge numbers of examples of mixes paired with their original separate tracks, learning to recognize the characteristic patterns of each type of sound.
What HTDemucs is
The model doing this separation is HTDemucs (Hybrid Transformer Demucs), a source-separation model trained specifically to split a musical mix into four categories: vocals, drums, bass and "other" (the rest of the instruments — guitars, keyboards, synths, etc.). It runs exported to ONNX directly in your browser, with the song never sent to any server — especially relevant given this is copyrighted music, which never leaves your device.
The four resulting tracks
- Vocals: the isolated lead vocal, the basis for making an acapella or for removing it to get the instrumental.
- Drums: percussion only, useful for studying a rhythmic pattern or sampling a break.
- Bass: the isolated bassline.
- Other: everything else — guitars, keyboards, strings, synths — anything that doesn't fit the previous three categories.
The four-track mixer
Once separated, the four tracks play back synchronized sample-accurately from the same time reference, with independent mute and solo controls per track. This lets you, for example, listen to just the drums and bass together to study how they interact, or mute only the vocals to get a full instrumental ready to sing over.
Real use cases
- Karaoke: mute the vocal track and keep the rest, without buying an official instrumental version (which many songs don't even have available).
- Sampling and music production: pull just the drums or bass from a song to use as the basis for a new project.
- Music study: isolate a specific instrument to transcribe it or learn its exact part, without the rest of the mix getting in the way.
- Remixes and mashups: separate a song's vocals to combine them with another song's instrumental.
Why it's a heavy task, and why it's worth it
Source separation is one of the most computationally demanding AI tasks that exists in the browser — the model has to process the whole song, analyzing complex frequency patterns to tell each instrument apart. It can take longer than other AI tools on this site. In exchange, the song never leaves your device: no server processes or temporarily stores the copyrighted audio file.
Frequently asked questions
Is the separation perfect? Not always — depending on the complexity of the original mix, slight traces of one instrument can bleed into another track, especially in very dense mixes. For most popular songs the result is surprisingly clean.
Is the song uploaded to a server? No, the whole process — from decoding to AI separation — happens in your browser.
Can I download the tracks separately? Yes, each of the four tracks (vocals, drums, bass, other) can be downloaded independently.
How long does it take to process a song? It depends on the audio's length and your device's power; it's one of the most demanding tasks on this site precisely because of the model's complexity.
Split any song into vocals, drums, bass and other instruments with the AI stem separator, 100% in your browser, with your music never uploaded anywhere.