The 2026 Remix Workflow: How AI Stem Separation Changed Everything
You used to need the original session files. Now a finished stereo track can be split into vocals, drums, bass and the rest. Here is what neural separation…
· Stem Separation
For most of recorded music history, a finished stereo mix was a one-way door. Once the vocal, drums and bass were summed together, you could not get them back apart. Remixers needed the multitracks, and multitracks mostly lived on a label hard drive nobody would share.
Neural source separation changed that. It is one of the two places Level uses AI (the other is Vocal Remover), and it is worth understanding what it is actually doing.
How it works, in plain words
A separation model has been trained on thousands of songs where the individual parts were available. It learned what a voice, a kick drum or a bass line "looks like" in the signal. When you give it a new mixed track, it estimates which part of the sound belongs to which instrument and rebuilds each one as its own file.
Two things follow from that. First, the model only separates what is already in your file — it does not generate new music, voices or lyrics. Second, it is an estimate. Where two instruments share the same frequencies at the same moment, some bleed is unavoidable.
What you get
Level splits a track into the classic four stems:
- Vocals — lead and backing voices.
- Drums — kick, snare, hats and percussion.
- Bass — bass guitar, synth bass and most of the sub.
- Other — everything else: keys, guitars, pads, FX.
Where it shines and where it struggles
- Shines: modern pop, hip-hop, EDM and rock, where instruments are well defined and mixed clearly.
- Shines: pulling a vocal to rebuild a track in a different genre, or drums to study a groove.
- Struggles: heavy reverb, which gets split between stems and can sound "washy" once isolated.
- Struggles: instruments in the same range at the same time, such as a distorted guitar under a vocal.
- Struggles: low-quality sources. A 128 kbps MP3 has already thrown information away that no model can recover.
A practical remix workflow
- Start from the best source you have — WAV or FLAC beats MP3 every time.
- Separate once, then listen to each stem solo before you plan anything. Decide what is usable.
- Use the stems as layers, not as a finished mix. A little bleed disappears once new parts sit on top.
- High-pass the vocal stem gently (around 80–120 Hz) to remove any kick or bass leaking in.
- If the "other" stem is messy, rebuild those parts instead of fighting them.
- Load the stems straight into the Level Mixer Lab to test ideas against a second track before you open a DAW.
Quality settings
Separation runs on our servers, and higher quality means the model does more passes over your audio. Top plans can use Level Studio Quality, which spends more processing per track for cleaner edges. For most remix work the standard setting is already very good; try the higher one when a stem will be heard exposed.
Rights and privacy
Separating a track does not give you rights to it. Remixing a song you do not own for release still needs permission. On our side, your uploaded file is deleted within 24 hours of the job finishing, and your audio is never used to train any AI model. Full details are in our Privacy Policy.
Level Audio runs in your browser and needs JavaScript enabled.
Contact: support@levelaudio.co