Industries / Media, arts and recreation / Media and entertainment / Music recording
Mixing and mastering
Mixing blends the separate recorded tracks into one finished version, in stereo or in an immersive format; mastering then sets that mix's level, tone and loudness for each release format.
7
Milestones achieved, from 1979 to 2023.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1979
The console gets a memory
Solid State Logic launched the SL 4000 E mixing desk in 1979. Sound On Sound calls it the first to combine "total parameter recall" with level automation, tape control, and tone and dynamics controls on every channel. Engineers could record, edit and overdub their fader moves to floppy disk, which made mixes of 24 or more tracks manageable and made it practical to come back to a mix later.
LessonTotal Recall only recorded where the knobs were; an engineer still reset each one by hand, and the producer Bobby Owsinski notes a recalled mix "might not sound the same".
Credit
Solid State Logic, founded in 1969 by Colin Sanders
2008
Death Magnetic's loudness backlash
Metallica's album "Death Magnetic" was criticized for "extreme brickwall limiting" and audible distortion. A limiter caps the peaks of a track so the whole of it can be pushed louder; Sound On Sound says the CD's dynamic range, the gap between loud and quiet, averaged 3 dB.
Lessonfor its 2016 album the band held a blind shoot-out between four mastering engineers, and the digital release measured 6 to 7 dB.
Credit
Metallica, with producer Rick Rubin and mastering engineer Ted Jensen; Sound On Sound says blame was pointed at both, but "the inside story… was never revealed"
2014
Players turn loud masters down
The ITU, the UN agency for telecommunications standards, published BS.1770 in July 2006, a standard way to measure how loud a recording sounds, in LUFS (loudness units relative to digital full scale). In February 2014 Sound On Sound reported that Apple's Sound Check aimed at about -16 LUFS and that Spotify had normalized loudness "from day one": each track plays back at the same loudness, so a heavily squashed master gains nothing and "ends up sounding dull and feeble". Spotify now says it adjusts tracks to -14 LUFS under ITU 1770 and turns louder masters down.
Lessononce players set the level, mastering louder stopped paying; the engineer Bob Katz told an Audio Engineering Society convention that "the loudness wars are over".
Credit
ITU-R (the standard); Apple and Spotify (normalization)
2014
Mastering by upload
On 12 May 2014 MixGenius of Montreal launched LANDR, an online service that masters an uploaded track automatically. Outside reviewers tested it in 2023: Mike Levine in Mix found the plug-in version's AI settings "consistently excellent", and MusicRadar rated LANDR best of four such services, at 8.5 out of 10.
Lessonautomatic loudness still needs a spec check; Levine had to pull the plug-in's default true peak (the highest level after conversion to sound, about -0.2 dBFS) below -1, and MusicRadar found that eMastered did not measure true peak at all, which "could cause a master to be rejected".
Credit
MixGenius, now LANDR
2021
Atmos mixes reach phones
On 17 May 2021 Apple announced Spatial Audio with Dolby Atmos on Apple Music for June 2021, at no extra cost. Dolby Atmos places sounds around and above the listener, and Apple said thousands of tracks would be ready at launch, playing automatically on AirPods and Beats headphones with an H1 or W1 chip. In January 2024 Hypebot reported an Apple memo to labels: tracks with a Spatial Audio version would earn 10% more in royalties, counted on all streams of that track.
Lessona second, immersive mix became something labels are paid extra for.
Credit
Apple Music and Dolby
2021
Stem splitting gets a scoreboard
Source separation splits a finished mix back into its stems: vocals, drums, bass and the rest. In 2019 Alexandre Défossez and co-authors described Demucs, which works directly on the sound wave and averaged 6.3 on the MusDB test set, in SDR (signal-to-distortion ratio, in decibels; higher is closer to the real stem). Sony's Music Demixing Challenge 2021 then scored entries on a hidden 30-song set made by Sony Music Entertainment (Japan); it drew 1,541 submissions from 61 teams, scores rose about 1.5 dB over the contest, and Défossez's Hybrid Demucs won.
Credit
Yuki Mitsufuji and co-organizers (Sony); Alexandre Défossez
2023
Demixed demo wins a Grammy
For the Beatles' "Now and Then", Peter Jackson's company used MAL, a stem-separation tool it built for the series Get Back, to split John Lennon's 1970s cassette demo into piano, Lennon's voice, and everything else in the room. Spike Stent mixed the track and Miles Showell mastered it, and it won Best Rock Performance at the 2025 Grammys.
Credit
WingNut Films' audio team, led by Emile de la Rey; Giles Martin and Paul McCartney
Watch for
Undoing a limiter
iZotope says Ozone 12's Unlimiter, a machine-learning module, is "capable of reviving overly compressed audio" by putting back lost peaks and dynamic range, treating limiting as "a sample-by-sample gain curve that can be inverted". It credits Chang-Bin Jeon and Kyogu Lee's 2023 de-limiter paper, and says it "did not focus on severely brick-walled or distorted sources". Sound On Sound's December 2025 review found the effect "can be pretty subtle" on a busy, loud mix.
Achieved whenan outside party runs it on a public test set, measures how close the output comes to the original, unlimited mix, and publishes the numbers.
Claimed by
iZotope, September 2025
Neural upmix near human
S2X-Net turns stereo into 5.1 or 7.1 surround, and its authors say the output reaches "near human-made upmixing results" in their own listening test with 50 people.
Achieved whenan independent group repeats the test on music outside the authors' test set, including formats with overhead speakers, and publishes the scores.
Claimed by
Marian Negru, Bogdan Moroşanu and Ana Neacşu, in IEEE Transactions on Audio, Speech and Language Processing, October 2026
Mixing by plain request
DAWZY is an open-source assistant that uses GPT-5 to turn typed, spoken or hummed requests into actions that can be undone in the REAPER audio editor. In its authors' comparison it completed 4 of 9 trials, against none for Ableton-MCP; 21 users gave "Control" the lowest rating, 3.81 out of 5, and some tasks failed because the model "could not reliably convert between units".
Achieved whenan outside group runs it on a fixed, public list of mixing tasks and publishes a success rate.
Claimed by
Aaron C. Elkins and co-authors, San Diego State University, December 2025
Open
01
Score the unlimiters
A public scorecard for de-limiters. Score Ozone 12's Unlimiter, at default and maximum settings, the open De-limiter model, and no processing at all on every track of musdb-XL, a heavily limited copy of a standard test set, then run a small blind listening test.
The briefCredit · unclaimed02
An open Atmos test set
Ten songs from the MedleyDB multitrack collection mixed by hand in 7.1.4 (seven speakers at ear level, one subwoofer, four overhead), a fixed stereo fold-down of each, automatic upmixes of those fold-downs back to 7.1.4, and a blind test of the upmixes against the hand mixes, rendered for headphones.
The briefCredit · unclaimed03
AI masters against delivery specs
A repeatable audit. Send 20 mixes through each major automatic mastering service at default settings, measure loudness and true peak, then measure true peak again after encoding to 256 kbps AAC, the format Apple Music streams, and decoding. Count how many files break Spotify's guidance: true peak below -1 dBTP, or below -2 dBTP for masters louder than -14 LUFS.
The briefCredit · unclaimed
Open 01 · the brief
Score the unlimiters
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard for de-limiters. Score Ozone 12's Unlimiter, at default and maximum settings, the open De-limiter model, and no processing at all on every track of musdb-XL, a heavily limited copy of a standard test set, then run a small blind listening test.
Why it is still open
The only outside review is a listening impression ("pretty subtle"), and no measured comparison of the Unlimiter against an unlimited original was found. The research model reports 24.0 dB SI-SDR (a version of SDR that ignores overall level) on its own data, but nobody has tested the commercial product against it.
Built with what exists
MUSDB18-HQ (150 songs with stems, for educational use only); the musdb-XL gain files (CC BY 4.0), which rebuild the limited copy from MUSDB18-HQ; the De-limiter code and weights, which its paper says are public; Ozone 12 Advanced; an SI-SDR scoring script and a BS.1770 loudness meter. The advance behind it is machine-learning de-limiting, published in 2023 and shipped in a product in 2025.
First step
Download MUSDB18-HQ (Zenodo record 3338373) and the musdb-XL gain files (Zenodo record 7041331), rebuild the limited tracks, and score them unprocessed against the originals as the baseline.
Done looks like
A published table of SI-SDR (dB) and loudness (LUFS) for every musdb-XL track under each condition, with code, plus a blind test with at least 20 listeners on 10 clips and the tally.
Open 02 · the brief
An open Atmos test set
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Ten songs from the MedleyDB multitrack collection mixed by hand in 7.1.4 (seven speakers at ear level, one subwoofer, four overhead), a fixed stereo fold-down of each, automatic upmixes of those fold-downs back to 7.1.4, and a blind test of the upmixes against the hand mixes, rendered for headphones.
Why it is still open
The ImmersiveFlow authors wrote in January 2026 that "no publicly available dataset contains high-quality music recordings in the 7.1.4 format", so they tested on a private set of 100 songs. S2X-Net stops at 5.1 and 7.1, with no overhead channels. No published test of automatic Atmos upmixes against engineers' mixes of the same stems was found.
Built with what exists
MedleyDB audio (122 multitracks, CC BY-NC-SA 4.0, non-commercial research only, access on request); the commercial upmixers Halo Upmix (NUGEN Audio) and WavDSP UpMix; ImmersiveFlow, whose code its authors list on GitHub; webMUSHRA for the listening test; a KEMAR dummy-head filter set for headphone rendering, as in the ImmersiveFlow test. The advance behind it is generative models that produce 7.1.4 audio directly from stereo.
First step
Request MedleyDB access, pick ten songs, and mix one in 7.1.4 as a pilot.
Done looks like
10 hand-made 7.1.4 mixes and their fold-downs published under the same share-alike licence, plus a listening test with at least 20 listeners and every score per upmixer and song.
Open 03 · the brief
AI masters against delivery specs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A repeatable audit. Send 20 mixes through each major automatic mastering service at default settings, measure loudness and true peak, then measure true peak again after encoding to 256 kbps AAC, the format Apple Music streams, and decoding. Count how many files break Spotify's guidance: true peak below -1 dBTP, or below -2 dBTP for masters louder than -14 LUFS.
Why it is still open
The 2023 MusicRadar test of four services gave no loudness figures; its one technical finding was that eMastered does not measure true peak. Mix found LANDR's plug-in defaulting to about -0.2 dBFS. No published audit with measurements after encoding was found.
Built with what exists
The services LANDR, eMastered, CloudBounce and Masterchannel; Apple's AURoundTrip and afclip tools, now part of macOS, which make and decode 256 kbps AAC files and flag clipping; a BS.1770 meter; your own mixes, or openly licensed mixes used non-commercially.
First step
Make or collect five licence-cleared mixes, measure them before mastering as a baseline, and run them through one service.
Done looks like
A published spreadsheet of loudness and true peak before and after encoding for 20 mixes on each service, a pass and fail count per service, and the same audit repeated six months later.