Source Separation
Source separation splits a mix into parts, and quality depends on overlap and phase.
Related
A single stereo file may be all you have, yet you want to change only the vocal or the drums. Source separation pulls the mixture apart.
The problem
A mix presses every part into two channels, where frequency ranges, phase and reverb overlap. Separation must decide which source owns each frequency at each moment; a wrong decision leaves bleed.
How it works
A common approach predicts a time-frequency mask for vocals, drums, bass or other sources, then rebuilds the waveform. Vocals and drums differ in range and attack, so they separate well; two guitars or a string section overlap and pull each other along. Phase and reverb leave a metallic or hollow residue.
Practice and judgement
Decide the use first. Choose a vocal model for an a cappella, a multi-stem model for a remix. Match sample rates before separation, check bleed afterwards, clean with EQ and gating, and avoid heavy processing on separated stems. Keep an unprocessed copy.
What to listen for
Hear whether sibilance is dulled, drum transients softened and backing vocals left in the vocal stem. Then hear the room tail in solo; separated reverb sounds unnatural.
Common misunderstandings
Separation gives you stems; cleaner is always better; separated tracks can be remixed freely. Separation is an estimate, not a recovery, and artefacts grow with processing.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Sources
- 音源分离的时频掩码与深度模型方法，依公开论文核查，未抄文本
- 分离轨的伪影与串音评估，依音乐信息检索资料核查，未抄文本
Used for fact-checking only; all prose is original.