How much faster is mastering in 3.3?

The standard preset processes a four-minute track in about 24 seconds instead of 69 — nearly three times quicker. The speed-up changed nothing audible: loudness and peak figures match the previous build to within a tenth of a decibel.

Why did progress used to stall at 85–90%?

That stage runs the loudness convergence module. It performed up to three full limiter passes over the whole track, each at eight-times oversampling. That oversampling ratio is a requirement for measuring inter-sample peaks, not for the limiter itself, so passes are now four times cheaper and the redundant ones are skipped entirely.

What does 'presets hit their stated loudness' mean?

If a card says −9 LUFS, that is roughly what you get: the worst deviation is now about 1 LU versus 3.5 LU before. The target is also identical across the app, the public API and AI hints — Hip-Hop used to promise −13 LUFS while the API delivered −10.

Will my previous masters sound different?

Existing files are untouched. New masters from loud presets land closer to the stated figure, and hot sources such as Suno or Udio tracks no longer lose their midrange, because level is matched before compression rather than after.

Magic Master 3.3: nearly 3× faster, accurate loudness, all 23 presets

2026-07-25 · Magic Master

Version 3.3 is mostly about speed and accuracy. Nearly all of it grew out of user feedback and specific complaints.

Mastering is nearly three times faster

The standard preset now processes a four-minute track in 24 seconds instead of 69, and progress no longer freezes at 85–90%.

The delay was not caused by heavy processing but by how it was organised. That stage runs loudness convergence: add gain, limit peaks, measure, repeat until the target is met. Each pass ran at eight-times oversampling — the track was temporarily inflated eightfold and pushed through the limiter in full. Three passes on a long track meant a minute and a half of waiting.

Eight-times oversampling is a standards requirement for measuring inter-sample peaks, not for the limiter itself. We separated the two: the meter stays at 8×, the limiter runs at 4×, and a quick pre-check now skips the heavy pass entirely when the peak already sits within the ceiling. On quiet presets that removes both iterations.

The sound did not change: loudness and True Peak match the previous build to within a tenth of a decibel.

Presets hit the loudness they promise

Picking EDM at −9 LUFS used to return −11.5, and Metal missed by more than three decibels. The same limiter was at fault: an overly long release held gain at the track's minimum, so it behaved like plain peak normalisation rather than a limiter.

The worst deviation is now about 1 LU instead of 3.5, and eleven of fifteen tested presets land within 0.11 LU. Industry distortion tests (SMPTE RP120, aliasing, multitone) returned the same figures as before — the extra loudness came without added grit.

A quieter inconsistency disappeared along the way: the Hip-Hop card promised −13 LUFS while the same preset via the public API delivered −10. The value is now identical everywhere — app, API and AI hints.

Loud sources keep their midrange

When a track arrived already loud — exported with a limiter on the master bus, or generated by Suno and Udio — processing compressed it far deeper than intended. Compressor thresholds are absolute, and level matching happened after compression. You heard it as a scooped midrange, a boxy character, a harsh top end and a drifting stereo image.

Level is now matched before compression. On a brickwalled mix the midrange shift dropped from 3.3 dB to 0.8 dB. Quiet and normal-level tracks are untouched: staging only attenuates.

The full breakdown is in Your track sounds dull after mastering.

All 23 presets are in the app

Trap, Pop, Rock, Metal, R&B, Country, Afrobeats, K-Pop and Jazz worked in the engine and the public API, but there was no way to choose them — the cards simply did not exist. They are all available now, and to keep 23 cards from reading as a wall they are grouped into Genres, Voice, Character and Special. Your chosen group is remembered.

The AI assistant actually works

The assistant was down: our model provider retired the name configured in settings, and every request failed. From the outside it looked fine — the chat politely said "try again in a minute" and preset suggestions quietly fell back to simple rules.

It now reads the metrics of your specific track and explains its choice: "crest factor 6.4 dB and LRA 2.0 — the track is over-compressed, use AI Unlimiter". Auto-mastering receives the same data and got more accurate. A health check was added so this kind of outage cannot go unnoticed again.

Small things you notice daily

The English interface is fully translated — 86 elements were still in Russian: batch processing, PRO modules, the request form and result buttons.

It is easier to read. We raised the contrast of cards and muted text in both themes: cards used to blend into the background and muted text fell short of the accessibility standard.

The logo is now consistent across every page and no longer squashes on narrow screens.

Try it in the app — track analysis is free and shows loudness, True Peak and crest factor before any processing.

Try Magic Master

Upload a track — a finished master in seconds.

Open mastering →LUFS analyzer