demix

Demixing is the computational reversal of downmixing, the process that combines multiple audio channels or stems into a single mixed track. Where a stereo master might contain drums, vocals, bass, and guitars all blended together, demixing attempts to isolate and recover each element as a separate stem. This is done entirely in the digital domain using spectral analysis and machine learning models trained on isolated instrument recordings.

The technical challenge lies in the lossy nature of mixing. When you sum four tracks to stereo, phase information is collapsed, frequency overlap occurs, and dynamic range is compressed. Demixing cannot recover what was never preserved; instead, it makes educated guesses about what frequencies and timing likely belong to drums versus vocals by recognizing acoustic patterns. Modern demixing algorithms use neural networks trained on large datasets of isolated instruments, learning to predict which parts of a spectrogram correspond to which source.

Common applications and limitations

Demixing appears in music production workflows where the original multitrack session is unavailable but remixing or repurposing is required. Audio restoration engineers use it to isolate speech from background noise or separate accompaniment from a vocal performance. Transcription services employ it to clean source material before processing. The quality of results depends heavily on the complexity of the original mix; a heavily compressed, heavily EQ'd master with many overlapping instruments will demix far worse than a live performance with clear instrumental separation.

The term itself merges the prefix de-, meaning to reverse, with mix. It is distinct from unmixing, which implies complete and perfect separation (generally impossible), and from source separation, a broader field covering any extraction of components from a combined signal. Demixing is increasingly accessible through software plugins and cloud services, though results remain imperfect enough that professional workflows still depend on preserving original multitrack recordings whenever possible.