What music source separation actually is
Music source separation is the task of recovering the individual sources inside a mixture. A finished song is a sum: every instrument and voice was recorded or generated on its own track, then added together, with equalisation, compression and effects, into two stereo channels. The mixing engineer had all the original tracks. You only have the sum. Separation tries to run the sum backwards.
That is an under-determined problem, because many different sets of tracks could add up to the same stereo file. A model cannot solve it by arithmetic. It solves it by knowing what sources sound like. A neural network trained on thousands of songs where the individual tracks were known learns what a snare, a bass line or a sung vowel looks like in the mix, and it estimates how much of each sound belongs to each source.
The result is an estimate, not a recovery. That one sentence explains almost everything you will hear in the output: why drums and vocals come out well, why a shaker hidden behind a guitar may smear, and why the stems of a very busy mix never sound quite as clean as the original recording session.
How Demucs htdemucs works, at a high level
TuneSplit runs Demucs, the open-source separation model family from Meta AI Research, in its htdemucs form. The name stands for hybrid transformer Demucs. You do not need the maths to use it, but the three ideas in the name help you predict the results.
- Hybrid. The model looks at the sound two ways at once. One branch reads the raw waveform, which preserves timing and the sharp attack of a snare. The other branch reads a spectrogram, a picture of which frequencies are loud at each moment, which makes pitched sounds like vocals and bass easier to follow. The two views are combined.
- Transformer. In the middle of the network, attention layers let every moment of the audio compare itself with other moments. That helps with context, such as knowing that a sound at 0:12 is the same synth that appeared at 0:04.
- Demucs. The outer shape is a U-Net style encoder and decoder. The encoder compresses the audio into a compact representation, and the decoder expands it back into one waveform per source.
For each source the network predicts a soft mask or a waveform, and the stems are written out. The model is trained on multitrack datasets where each stem is known, and it is scored by how closely its output matches the real stem, a measure called signal-to-distortion ratio, or SDR. Higher is better. Reported scores for htdemucs on the standard MUSDB benchmark average in the region of 9 dB, which was a clear step up from earlier models, but numbers vary by training data and test set, so treat that as a rough guide rather than a promise about your own file.
Hear a real four-stem separation
The players below hold a real separation made by the same engine as the tool above. The source is Sunlight, an indie pop track we generated with TuneSplit, so there is no copyright question. You get the original mix, then each stem on its own. Play the original, then solo the stems one at a time.
The split covers the first 60 seconds of the song and took 88 seconds from upload to finished files in one measured run. That is one data point, not a guarantee. For a typical 3-minute track, the usual range is 20 to 60 seconds. When you listen, focus on the gaps between vocal phrases and between drum hits. A good result is close to silent in those gaps.
Phase cancellation, spectral EQ or AI: three ways to separate music
People have tried to pull songs apart for decades, and the older methods still turn up in free editors and plug-ins. They work in very different ways, so the results differ too. Here is how the three families compare.
| Approach | How it works | What it can do | Main weakness |
|---|---|---|---|
| Phase cancellation | Flips one stereo channel and sums it with the other, so anything panned dead centre cancels out | Mutes a centre-panned lead vocal in some older stereo recordings | Also cancels centre-panned kick, snare and bass, leaves a thin hollow mix, fails on mono files and modern wide vocals |
| Spectral EQ and filtering | Cuts or boosts frequency bands, or gates by pitch range, to reduce the part you do not want | Rough bass boost or cut, dulling a voice in a crowded band | Vocals and guitars share the same frequencies, so you remove part of everything, not one source |
| AI source separation | A trained neural network estimates each source from the waveform and the spectrogram | Vocals, drums, bass and other as clean separate files, on mono or stereo, regardless of pan position | An estimate, so you can hear faint bleed or watery artefacts on dense mixes |
Phase cancellation is the old karaoke-machine trick, and it only works when the singer was recorded dead centre and nothing else shares the centre, which in practice means almost nothing from the last 30 years. EQ cannot choose by source at all, only by frequency. AI separation is the first approach that decides by what a sound is, not where it sits, which is why it gave usable stems from a plain stereo file for the first time.
Artefacts and bleed: what you may hear, and why
Every separation tool leaves some trace of the work. Knowing the common signs helps you decide whether a result is good enough for your purpose.
| What you hear | Likely cause | What to try |
|---|---|---|
| Faint hi-hat or cymbal ticks in the vocal gaps | High-frequency percussion overlaps with breath and sibilance | Fade or gate the gaps; try a lossless source |
| Watery or phasey tone on the stem | The model blended its estimate across neighbouring frequency bins | Compare to the original; most noticeable on sustained notes |
| Reverb tail of a vocal sitting in the music stem | The tail overlaps with pads and room sound | Use a short fade or a de-reverb step afterwards |
| Kick drum partly in the bass stem | Both occupy the same low frequencies | Use the drums stem plus a low-pass on the bass if needed |
| Backing vocals in the vocals stem | Stacked voices are vocals to the model | Edit by hand if you want the lead only |
| A sound missing from every stem | The model could not decide where it belonged | Check the other stem; try another version of the song |
Bleed is not a bug in one tool. It is the price of estimating sources from a sum. The better the model, the smaller it gets, but it never drops to zero on a dense, heavily compressed mix.
What quality to expect by genre
Quality depends far more on the music than on the settings. These are rules of thumb from how the model behaves, not measured scores for each genre.
| Genre | Usually works well | Usually harder |
|---|---|---|
| Pop and R&B | Vocals, drums and bass are clearly distinct and mixed with space between them | Heavy vocal stacks and layered synth pads |
| Hip-hop and trap | Vocals and 808s separate well; drums are crisp | Sub-bass and kick sit in the same range |
| Rock and metal | Drums and bass separate well | Distorted guitars are dense and overlap vocals and cymbals |
| Electronic and dance | Kick, bass and lead vocal stand out | Sidechained pads and wide synths blur into 'other' |
| Acoustic and singer-songwriter | Voice and guitar are sparse, so often the cleanest results | Close-mic room sound can bleed between stems |
| Jazz and classical | Drums, bass and piano can work on small ensembles | Many similar instruments and reverberant halls blur sources |
| Live recordings | Drums and vocals can still be useful | Crowd noise and room sound leak into every stem |
2 stems or 4 stems: which split to choose
The tool above offers two layouts. Pick by what you plan to do with the files, not by which sounds more advanced.
| Layout | Files you get | Best for |
|---|---|---|
| 2 stems | Vocals and music (everything except the voice) | Karaoke, acapellas, singing along, removing a voice from a clip |
| 4 stems | Vocals, drums, bass and other | Remixing, learning a bass line, drum sampling, mixing study, practice with one part muted |
- Start with 2 stems if all you need is a voice or a backing track. It is the simplest result to judge and use.
- Use 4 stems if you want to mute or rebalance one instrument, such as playing bass over a track that has no bass.
- Remember the other stem is a catch-all: guitars, keys, synths, strings and effects share one file.
- Keep the original mix next to the stems and combine them in your editor to check nothing important was lost.
Both layouts count as one file against your daily allowance, so you can try the 4-stem layout and still have free files left to try the 2-stem one on the same song.
Music Separator facts at a glance
| Item | Detail |
|---|---|
| Price (free tier) | 3 files a day, free; after that 30 credits per started 5 minutes |
| Download format | Every stem as MP3, free |
| Sign-up | Not needed for the free files |
| File deletion | Uploads and results are deleted within 24 hours |
| Model | Demucs htdemucs (hybrid transformer) |
| Speed | Usually 20 to 60 seconds for a 3-minute song |
| Limits | Up to 10 minutes per file; MP3, WAV, M4A, FLAC, OGG or MP4 in |
Credit prices are on the pricing page. The free daily files never need credits.
Getting the best result from a separator
- Use the best source you have. A lossless FLAC or WAV beats a low-bitrate MP3, because compression removes the fine detail the model uses.
- Prefer the studio version. Live takes and bootleg recordings carry room sound that bleeds into every stem.
- Judge on headphones, not a laptop speaker, especially for the bass stem.
- Listen to the quiet gaps first. That is where leftover sound shows up.
- Separate a clean section if you only need a loop. Short, sparse passages often give the best results.
- Do not stack separations. Running stems through the tool again adds artefacts rather than removing them.
If you are separating your own recording, export a dry or lightly processed bounce when you can. Heavy mastering compression glues sources together and makes the job harder.
Your upload, your rights
Only upload audio you own or are licensed to use. That includes your own songs, session bounces, public domain recordings and audio from collaborators who agree. Separating a track does not give you the right to publish its parts, and posting an acapella or instrumental of another artist's song without permission can infringe their rights.
Free songs and stems are for personal use. Your upload is held in private storage, used only to produce your result, deleted within 24 hours and never used for training. TuneSplit has no voice cloning and no AI covers of real artists. Rights holders can use the DMCA contact on the Terms page.
Frequently asked questions
What is a music separator?
A music separator is a tool that splits a finished song into its separate parts, such as vocals, drums, bass and other instruments. TuneSplit uses the Demucs htdemucs neural network to estimate each source from the stereo mix, and returns each one as its own audio file you can play and download.
Is this music separator free?
Yes, within limits. You get 3 files a day, up to 10 minutes each, and every stem downloads as an MP3. You do not need to sign up for the free files. After the third file in a day, each started 5 minutes of audio costs 30 credits.
Is an audio separator the same as a stem splitter?
Mostly yes. Audio separator, music separator, stem splitter and source separation all describe the same job: pulling the layers out of a mixed track. The names differ by who is searching. The TuneSplit stem splitter page covers practical uses, while this page explains how the separation works.
How does AI music separation work?
A neural network is trained on songs where each instrument's track is known. It learns what vocals, drums and bass look like inside a mix. Given a new song, it estimates the share of each sound that belongs to each source and outputs one waveform per source. It is an estimate, not an exact recovery.
What does htdemucs mean?
It stands for hybrid transformer Demucs, a model from Meta AI Research. Hybrid means it reads both the waveform and a spectrogram. Transformer refers to attention layers that look across the whole segment for context. Demucs is the model family, a U-Net style encoder and decoder that outputs one track per source.
Why do my stems still have some bleed?
Because separation estimates sources from a sum and does not recover the original tracks. Sounds that overlap in frequency, like hi-hats and breath or kick and bass, are hardest to divide. Dense mixes, heavy compression and low-bitrate files make it worse. Starting from a lossless studio version helps.
Is the phase cancellation trick better than AI?
No, except in rare cases. Phase cancellation only removes sounds panned exactly to the centre, so it also removes the kick, snare and bass and leaves a hollow result. It does nothing on mono files. An AI model decides by what a sound is and works on any stereo or mono file.
Should I choose 2 stems or 4 stems?
Choose 2 stems if you just need a voice or a backing track. Choose 4 stems if you want to remix, mute or rebalance one instrument, learn a bass line or sample drums. Each file counts once against your daily allowance, whichever layout you pick, so trying both costs two files.
How good is the quality for my genre?
Pop, hip-hop and acoustic music usually separate cleanly, because the parts are distinct. Dense distorted guitars, orchestral recordings and live tracks are harder and show more bleed. Use the genre table on this page as a guide, and listen to the result before you download it.
How long does separation take?
A 3-minute song usually takes 20 to 60 seconds. As a reference, the 60-second demo on this page took 88 seconds in one measured run. Longer files take longer, and busy periods can add waiting time, so treat the usual range as typical, not guaranteed.
Are my files private?
Uploads go to private storage, are used only to produce your result, are deleted within 24 hours and are never used to train any model. You should only upload audio you own or have the rights to use, because you remain responsible for what you do with the output.
Can I use the separated stems in a release?
Only if you hold the rights to the underlying recording. Stems from your own music are yours to use. Stems from a commercial release are for private study unless you have permission. Free results on TuneSplit are for personal use, and a paid plan adds a commercial licence for what you generate there.