BRENNEN

Industries / Media, arts and recreation / Media and entertainment / Music recording

Writing and producing

Writing and producing covers making a song and shaping its recording: writing melodies and lyrics, arranging parts, building beats, sampling older records, and the tools for this work, from drum machines and samplers to pitch correction and AI song generators.

7

Milestones achieved, from 1980 to 2025.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1980

    Drums from a box

    Roland's TR-808 drum machine made its drum sounds with analog circuits. Within four years it was on records made in Osaka, London, Bombay, Belgium and New York, by Roland's account. On "It's Yours" (1984) Rick Rubin lengthened its kick drum until, in Roland's words, it "became the heaviest instrument in the room". Early users included Yellow Magic Orchestra, Ryuichi Sakamoto and Marvin Gaye.

    LessonRoland says that in Japan the machine "was being judged a commercial failure", and Roland had stopped making it by 1984, just as the records made with it were spreading.

    articles.roland.com

    Credit

    Roland Corporation, founded by Ikutaro Kakehashi in 1972

  2. 1983

    One language for instruments

    MIDI (Musical Instrument Digital Interface) is a shared standard for sending notes and control messages down a cable, so keyboards, drum machines and computers from different makers can play together. At the 1983 NAMM trade show a Sequential Prophet-600 was linked to a Roland Jupiter-6, and an early MIDI 1.0 specification is dated August 1983.

    Lessonthe MIDI Association says the story of an overnight success "was very far from the reality" of the work between January 1983 and May 1985, when the MIDI Manufacturers Association was formally set up.

    midi.orgroland.com

    Credit

    Ikutaro Kakehashi (Roland) and Dave Smith (Sequential Circuits), who received a Technical Grammy for MIDI on 9 February 2013

  3. 1988

    Sampler, sequencer and pads

    The Akai MPC60 went on sale in 1988. It combined a sampler (13.1 seconds of mono audio, or 26.2 seconds with extra memory), a 4-by-4 grid of pads to play the samples, and a sequencer, a recorder of notes, that could also drive four other MIDI instruments. A whole song could be programmed on one box. Sound On Sound lists Dr Dre, J Dilla, DJ Premier and DJ Shadow among its users.

    LessonSound On Sound says J Dilla built his loose drum style by switching the machine's timing correction off, so a tool's defaults are not its limits.

    soundonsound.com

    Credit

    Roger Linn, working with Akai Professional, whose team included Dave Cockerell

  4. 1991

    Samples need permission

    In Grand Upright v. Warner (US federal court, New York, 1991), Judge Kevin Duffy opened with "Thou shalt not steal", blocked Biz Markie's use of Gilbert O'Sullivan's "Alone Again (Naturally)" and referred the case to the US Attorney. In Bridgeport v. Dimension Films (Sixth Circuit, 2005), the court held that even a two-second guitar sample, pitched down and looped, needed a licence: "Get a license or do not sample." In 2016 the Ninth Circuit disagreed in VMG Salsoul v. Ciccone, holding that a 0.23-second horn hit in Madonna's "Vogue" was too small to count, so US appeals courts are split.

    Lessonproducers clear samples or replay the part, and Bridgeport itself said an artist "is free to duplicate the sound of that 'riff' in the studio".

    blogs.law.gwu.edublogs.law.gwu.edublogs.law.gwu.edu

    Credit

    Judge Kevin Duffy; the Sixth and Ninth Circuit courts

  5. 1998

    Pitch correction as a sound

    Antares released its Auto-Tune plug-in in 1997 to fix out-of-tune vocals. On Cher's "Believe" (1998), producers Mark Taylor and Brian Rawling set its retune speed to zero, which snapped every note to the nearest pitch and made a robotic sound now called the Cher effect.

    Lessonin 1999 Taylor told Sound On Sound he had used a Digitech Talker pedal, and he later said on The South Bank Show that it was Auto-Tune, so a studio trick can be hidden for a while but not kept.

    mixonline.comantarestech.commusicradar.com

    Credit

    Andy Hildebrand (Antares); Mark Taylor and Brian Rawling

  6. 2020

    Courts narrow song-copying claims

    In March 2015 a Los Angeles jury found that "Blurred Lines" copied Marvin Gaye's "Got to Give It Up" and awarded US$7.4 million, by the Music Copyright Infringement Resource's (MCIR) account, and the Ninth Circuit upheld that in 2018. On 9 March 2020 the full Ninth Circuit restored Led Zeppelin's win over "Stairway to Heaven". It held that protection covered only what was in the sheet music deposited for registration, and it dropped the "inverse ratio" rule, under which proof that the defendant had heard a song lowered the similarity a plaintiff had to show. In May 2023 a unanimous jury cleared Ed Sheeran's "Thinking Out Loud", and on 16 June 2025 the Supreme Court declined a separate appeal, leaving rulings that the shared elements were too common to protect.

    Lessonafter the 2015 verdict, courts moved back toward treating chord progressions and stock rhythms as free building blocks.

    blogs.law.gwu.edublogs.law.gwu.edublogs.law.gwu.eduksl.com

    Credit

    the Ninth Circuit (Skidmore v. Led Zeppelin); the federal court in New York and its jury; the Second Circuit

  7. 2025

    Only humans author songs

    On 29 January 2025 the US Copyright Office said AI output is protected "only where a human author has determined sufficient expressive elements", and that "the mere provision of prompts" is not enough, while a person's arrangement of or changes to AI output can qualify. Courts kept the human-author rule: in Thaler v. Perlmutter, about an AI-made picture, the D.C. Circuit ruled in 2025 and the Supreme Court declined review on 2 March 2026.

    Lessonin the US, a writer using an AI generator owns only the parts they wrote, chose or arranged.

    copyright.govmayerbrown.com

    Credit

    the US Copyright Office; the D.C. Circuit

Watch for

  1. Licensed AI song models

    In its November 2025 settlement with Warner Music Group, Suno promised "new, more advanced and licensed models". On 9 September 2026 it launched v6, v6-wild and v6-mini and began retiring its older models. Suno says v6 was trained from scratch on licensed material from Warner Music Group, BMG and Believe plus user data; Warner's chief executive Robert Kyncl says the old model was trained "without permission". The Next Web notes that Suno names no catalogues and gives no amounts or training method, and Universal and Sony have not joined.

    Achieved whensomeone outside these companies, such as a court-supervised disclosure, an auditor or a certifier of training data such as Fairly Trained, confirms what v6 was trained on.

    thenextweb.commusicweek.combillboard.com

    Claimed by

    Suno, with Warner Music Group, BMG and Believe, September 2026

  2. Memorized songs ruled infringing

    On 31 July 2026 the Regional Court Munich I ruled for GEMA, the German collecting society, against Suno (case 42 O 763/25). It found that six songs, including "Atemlos durch die Nacht" and "Daddy Cool", were reproducibly contained, or "memorized", in Suno's v3.5 and v4 models, that Germany's exception for text and data mining did not cover those copies, and that Suno, not its users, was responsible for the outputs. The Next Web calls it the first such decision in Europe.

    Achieved whenthe judgment becomes final; the court says it is not yet final.

    justiz.bayern.de

    Claimed by

    Regional Court Munich I, first instance, July 2026

  3. Recordings into lead sheets

    A lead sheet is a one-page score with the melody, chord symbols and lyrics. SheetSage2, from Junyan Jiang, Ruibin Yuan, Gus Xia, Yann LeCun and others, turns recordings into lead sheets and timed notes on melody, chords, beat, key and song structure. Its authors say one model beats the earlier systems they list on 12 of 15 benchmark-and-measure pairs, and they report a vocal-melody score (F1, where 100 is perfect) of 82.51 on the RWC-Pop test set, against 62.71 for SheetSage1. The weights are public under a non-commercial licence.

    Achieved whenan independent group reproduces these scores, or tests the model on new material, and publishes the results.

    arxiv.orghuggingface.co

    Claimed by

    the SheetSage2 authors, October 2026

Open

  1. 01

    Demo to lead sheet, scored

    A pipeline that takes a songwriter's rough phone demo, voice with guitar or piano, and writes a lead sheet with melody, chord symbols and lyrics lined up bar by bar, plus a public scorecard that compares each output with a lead sheet written by hand.

    The brief
    Credit · unclaimed
  2. 02

    AI melody recall test

    A repeatable test that prompts AI song generators with the lyrics, title and style of public-domain songs, the same kind of prompt GEMA used in the Munich case, and measures how close each generated vocal melody comes to the original.

    The brief
    Credit · unclaimed

Open 01 · the brief

Demo to lead sheet, scored

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A pipeline that takes a songwriter's rough phone demo, voice with guitar or piano, and writes a lead sheet with melody, chord symbols and lyrics lined up bar by bar, plus a public scorecard that compares each output with a lead sheet written by hand.

Why it is still open

SheetSage2's numbers come from its own authors, on research test sets of finished recordings, not rough demos, and its model card lists melody, chords, beats, key and structure but not lyrics. The lyrics benchmark Jam-ALT measures finished songs from the Jamendo catalogue and has no public leaderboard. No published result was found for ElevenLabs Scribe v3 on sung lyrics.

huggingface.coaudioshake.github.io

Built with what exists

SheetSage2 (open weights, CC BY-NC 4.0, so non-commercial use only), the model that makes this possible now; ElevenLabs Scribe v3 (see the Note "Fewer word errors") or OpenAI's open Whisper model for lyrics; Jam-ALT's alt-eval package for the lyric error score. For an optional public-domain set: US sound recordings from 1925, in the public domain since 1 January 2026 by Duke's Center for the Study of the Public Domain, scored against period sheet music from Johns Hopkins' Levy Sheet Music Collection.

  1. First step

    Record five demos of songs you wrote, write their lead sheets by hand, then run SheetSage2 and Scribe v3 on one and score it.

  2. Done looks like

    20 demos of original songs, the builder's own or volunteers' with written consent, each with a hand-written reference lead sheet, and a published table of melody-note F1, chord-symbol accuracy and lyric word error rate for each song, with the code.

Open 02 · the brief

AI melody recall test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A repeatable test that prompts AI song generators with the lyrics, title and style of public-domain songs, the same kind of prompt GEMA used in the Munich case, and measures how close each generated vocal melody comes to the original.

Why it is still open

The Munich court found memorization only in Suno's v3.5 and v4 models and only for six songs. Suno's v6 was trained on new data whose make-up it has not published, and no published recall test of v6 was found. The nearest work, Guilherme Coelho's November 2025 paper, measures how close Udio and Suno outputs come to particular artists' styles, not recall of particular melodies.

justiz.bayern.dearxiv.org

Built with what exists

Suno v6 on a paid plan (Billboard reports that free accounts have a limit of seven downloads); SheetSage2 to pull out the generated vocal melody; the open melodic-similarity code and 40-case set of real copyright rulings from the comp-music-lab "music-copyright-expanded" repository on GitHub, to set a similarity threshold; songs on Duke's list of 1930 compositions that entered the US public domain in 2026, such as "Georgia on My Mind" and "I Got Rhythm" (check their status outside the US).

  1. First step

    Rerun the similarity code on the 40-case set to reproduce a threshold, then prompt three songs and score them.

  2. Done looks like

    20 public-domain songs, three generations each, and a published table of each of the 60 outputs' similarity scores and the count above the threshold, with 20 mismatched pairs as a baseline. Publish the notation and scores, not the generated audio, and state that the 40-case study found algorithms matched court outcomes at most 75% of the time.