MixSiftTwo different things you may hear
Bleed means you can still hear material from one intended part in another estimated stem: for example, a voice remains in the accompaniment after Vocals is muted. An artifact is a change introduced by separation, such as an unstable tail or altered attack. These are descriptions of possible symptoms, not measurements of a MixSift result.
A finished recording contains overlapping sources. MixSift estimates four groups rather than retrieving original isolated recordings. Results depend on the source; a group name does not certify a clean recording.
Check the cause before changing the balance
Compare the passage with the original recording at a comfortable listening volume. A sound already in the source is not evidence that separation created it.
Solo the relevant stem and briefly compare the other groups. Check Mute first if a soloed stem is silent, then clear Solo before returning to the full backing.
Ask whether the symptom matters for your use. A cue that helps private practice may be unsuitable for a backing you intend to present publicly.
If reducing another stem makes the remnant less noticeable, listen for useful music lost at the same time. Keep the original project and avoid treating that tradeoff as a repair of the separation.
What the mixer cannot fix
Mute excludes the selected stem, not every occurrence of that instrument throughout all four files. Level changes the contribution of a whole group. Neither control can selectively erase a vocal syllable that shares Other with useful accompaniment, or recreate an attack that was not retained.
A genuinely better source copy can be worth comparing, but there is no guaranteed clean result or universal setting. Converting a compressed file to WAV does not restore discarded source detail. See how to choose an input file.
Describe the issue safely
For support, begin with the app version/build, macOS version, chip, action attempted and any safe support code. Describe whether the issue is a failure to open/export or an audible quality concern. Do not send audio, raw logs, full paths, receipts, private project names or account identifiers. Redact screenshots before sharing.
No audio was processed or auditioned for this article. It cannot certify a particular recording’s quality. Support’s separation-quality section retains the same source-dependent caveat.