Establish a balance before asking for analysis
Begin with clip gain, faders, panning and intentional routing. If the vocal is simply too quiet, a complicated chain may solve a problem that a fader move would address more transparently. Build a static balance that communicates the song, then write down the remaining difficulty. “The chorus loses vocal intelligibility when the guitars enter” is a testable problem. “Make it professional” is not.
Assistants work from the signal and context they receive. A tool listening only to a bass track cannot fully judge its relationship with the kick unless that relationship is provided through an appropriate analysis or routing feature. Check what the tool hears, how long it listens and whether its recommendations are based on an isolated track, a reference or the full mix.
Use a proposal as a starting position
iZotope Neutron and Ozone, sonible smart:EQ and comparable products illustrate different forms of assisted processing. Their capabilities vary by edition and version. They may propose tonal correction, dynamics settings or mastering treatment; they do not share a single method, and the presence of an “assistant” does not tell you how a recommendation was produced.
Save the proposed setting, then compare it with a simpler manual alternative. If both solve the masking problem, prefer the one that preserves the performance and remains easy to revise. Avoid treating an attractive analysis display as proof. A broad reduction in the midrange may clean up a soloed guitar while weakening the aggression that the arrangement needs.
Level-match before choosing
Gain differences bias comparison. Adjust the processed version so that switching between it and the original does not create an obvious loudness jump. Short-term loudness readings can help, but listening remains necessary because changed frequency balance alters perceived level. Compare the same passage, not a verse in one version and a chorus in another.
Turn off the screen briefly. Ask whether the vocal is easier to follow, whether the low end has a clearer rhythmic shape and whether the snare still carries the intended impact. Listen quietly as well as at your normal working level. A setting that feels exciting only when loud may be exaggerating brightness or density rather than improving the balance.
Comparison, not a loudness contest
Illustrative equal-length bars, not audio measurements. Match the perceived level, then judge clarity, movement and tone.
Treat separated stems as imperfect material
Source separation estimates components from a combined recording. Research tools such as Demucs and Spleeter demonstrate this approach, but no separation method can guarantee recovery of the original multitracks. Shared frequencies, ambience and overlapping transients can produce leakage, metallic tails or unstable stereo information. A separated vocal may be adequate for a rehearsal reference and still unsuitable for an exposed release edit.
Audition transitions, breath sounds, cymbal decays and reverb tails. These often reveal damage more clearly than the loudest section. Compare in mono and avoid aggressive processing that exaggerates watery artefacts. Keep the original file available so you can test whether a blend, a narrower edit or a different arrangement solves the problem more naturally. Permission to possess a recording does not automatically grant permission to remix or release it.
Keep meters in their proper role
ITU-R BS.1770 describes algorithms for measuring programme loudness and true-peak level. EBU R 128 provides a broadcast loudness recommendation built around related measurement practices. These are useful technical references, not instructions to force every music master to one artistic outcome. A streaming service’s playback normalisation is also not a universal mastering target.
Use meters to find clipping risks, unexpected changes and delivery mismatches. Use listening to decide how much transient movement and density the song needs. If a limiter removes the groove, meeting a convenient number does not rescue the result. Keep headroom during mixing, and make final loudness choices against the actual destination’s current requirements.
Finish with a change you can explain
For every accepted suggestion, write a sentence about its audible purpose. “Reduced low-mid masking under the lead vocal” is more useful than “assistant preset two.” Print a reference mix, preserve the unprocessed route and review the result after a break. If you cannot reliably prefer the change, consider removing it.
Consult ITU-R BS.1770 and EBU R 128 for measurement context. When the mix works, the next task is a documented session handoff, not another round of automated optimisation.