How to split a song into stems: vocals and instrumental
Sometimes you need only the vocal from a finished track (for a remix or karaoke) or only the instrumental. Stem separation pulls those parts out of the final mix — done by an AI model trained to split audio by source.
Why you need stems
Karaoke and backing track — remove the vocal, keep the instrumental.
Acapella for a remix — isolate the clean vocal.
Replace a part — rewrite the instrumental under the same vocal.
Sampling — pull a fragment without the extra layers.
Why Suno own stems disappoint
Suno built-in stem export does not separate the finished mix — it essentially re-generates each channel referencing the original. That adds extra artifacts and ghosts of other parts, and stem-replacement attempts often fail. It is more reliable to take the final file and separate it with a dedicated source-separation model.
How AI separation works
Models like Demucs are trained on thousands of tracks to tell sources apart in a mix — voice, drums, bass, the rest — and reconstruct them separately. For most tasks two stems are enough: vocal and instrumental.
Split stems in the Lab
Our Lab has a Split Stems tool: upload a finished track and get a clean vocal and instrumental via a strong source-separation model. It costs 20 𝄞 and is included with catalog / Lab access.
Tip: once you have split the vocal and instrumental and processed them separately, mix them back together with the Combine Vocal tool.
FAQ
What do I get as output?
Two stems: a clean vocal and the instrumental (backing track). Use them for karaoke, a remix, replacing a part or sampling.
Why not split stems inside Suno?
Suno built-in export effectively re-generates each channel referencing the original and adds artifacts, so stem replacement often fails. Separating the finished file with a dedicated engine is more reliable.
Does it work with Suno tracks?
Yes. Separation works on any finished mix — from Suno, from a DAW or downloaded, in MP3 or WAV.
How much does it cost?
Stem separation is available in our Lab for 20 𝄞 (included with catalog / Lab access). The engine is a strong source-separation model.