Retrospective systems entries / 100 entries
Model papers, dataset records, industry economics, provenance standards, evaluation methods and the older precedents behind AI-assisted music, each read as a system with its craft and rights consequences.
Historical event dates and source dates are separate from the preparation date of this local edition. Every entry is a retrospective draft prepared 16 September 2026; none was published on its historical date.
100 entries
Hiller and Isaacson's own chronology names a mixed student quartet, not the full Walden String Quartet, at the piece's first hearing.
Mathews's own 1963 paper and a 2007 oral history document why Bell Labs' MUSIC programs made a computer a general synthesis instrument.
The Library of Congress and Bell Labs give different years and machines for the demonstration behind computer-synthesized singing.
The Xenakis research centre's own history dates UPIC's first prototype to 1977 and traces four decades of institutional continuations.
MIT Media Lab's own project page traces Hyperinstruments from 1986 virtuoso tools to systems for non-professional players after 1992.
Lewis's own writing and interviews describe Voyager's player-and-listener design as an independent improvising partner, not an instrument.
PG Music's own history and a 2017 review trace the product from 1990 MIDI-only output to recorded RealTracks added in 2006 and 2007.
Essl's own site describes a 1992 generative piano system whose modular rules keep producing new, non-reproducible output decades later.
Cope's own account of his 1996 book names deconstruction, signature extraction and recombination as EMI's three-step method.
Celemony's own timeline and a November 2001 review confirm the original release edited only monophonic audio, years before Direct Note Access.
Yamaha's own VOCALOID history and a 2007 engineering paper date the engine's 2003-2004 debut and name its synthesis method.
Wolfram's own site and a 2011 blog post date WolframTones to September 2005 and describe cellular automata, not a trained model, as its method.
Georgia Tech's own project and publication pages date Shimon to 2009 and trace its growth into composition and singing.
Company records and a contemporaneous report date Jukedeck's 2012 founding and an unconfirmed 2019 move to ByteDance.
Shutterstock's 2020 release and Amper's own 2019 site describe a licensed sample library and an assets-only acquisition.
Google and Magenta's own posts date A.I. Duet's public release to February 2017, after a December 2016 conference demo.
DeepMind's 2016 paper and blog report specific listening-test gains from raw-audio generation, not any later product's training data.
Magenta's 2018 paper and project page describe a hierarchical model for interpolating and sampling musical scores, not audio.
OpenAI's 2019 post names MuseNet's MIDI training sources and Sparse Transformer architecture, distinct from the later Jukebox system.
Harmonai's GitHub and Hugging Face pages document an MIT-licensed diffusion model with undisclosed training audio.
OpenAI's own paper and blog describe a crawled, unlicensed training corpus and hours-long render times.
Google's paper pairs semantic and acoustic tokens and flags a detection safeguard only for its speech results.
Google's paper discloses MusicCaps in full while leaving its 280,000-hour training corpus unattributed.
Google's paper documents a source-separation training method without naming the million-track corpus it drew on.
Meta's paper states a 20,000-hour licensed training pool by source, though the released weights stay non-commercial.
Meta's masked-decoding paper cuts latency by up to tenfold while its own hours figure for licensed data drifts across documents.
Stability AI's paper lists exact Freesound and FMA recording counts, distinct from the commercial Stable Audio licence.
A Stable Diffusion fine-tune on spectrogram images, published under MIT licence, now survives only in its archive and repository.
Suno's own repository calls Bark a separate text-to-speech-plus system and discloses no training data for it.
The open YuE lineage adds an editable melody-and-chord plan, with its Suno comparison flagged as its own unproven benchmark.
DeepMind's 2023 announcement of Lyria and Dream Track named nine partner artists and a separate producer-tested toolkit, distinct from later model versions.
ElevenLabs' own launch post and docs describe Music v2's section-by-section composition and inpainting, apart from its commercial licence terms.
Magenta's MAESTRO dataset pairs about 200 hours of competition piano audio with millisecond-aligned MIDI under a non-commercial share-alike licence.
Colin Raffel's Lakh MIDI Dataset licenses his own compilation of 176,581 scraped files, not the copyright status of each transcription inside it.
The 2016 FMA paper packaged over 100,000 Creative-Commons-tagged tracks from the Free Music Archive, though CC terms vary track by track.
Mozilla's Common Voice project crowdsources multilingual speech donations, validated by community vote and released under a CC0 public-domain licence.
The MTG-Jamendo dataset's own GitHub documentation restricts its 55,000-plus Creative-Commons tracks to non-commercial research and academic use.
Magenta's 2017 NSynth paper pairs a WaveNet-style autoencoder with a dataset of about 305,000 annotated single notes from sampled instruments.
Google's MusicLM paper describes MusicCaps as 5,521 AudioSet clips captioned by ten professional musicians for evaluation, not large-scale training.
Google's AudioSet documents 2 million labelled 10-second clips as YouTube references and precomputed features, not redistributed audio files.
The Copyright Office's Part 1 report recommends a new federal right against unauthorized digital replicas but leaves existing law otherwise unchanged.
The Copyright Office's Part 3 report rejects a blanket rule for AI training, weighing transformativeness against market harm case by case.
Music publishers' 2023 copyright suit over Claude and song lyrics moved from Tennessee to California and remains active with no ruling on the merits.
The AI Office's mandatory template sets a common minimum for what general-purpose AI providers must disclose about training content.
A 2024 consultation's preferred opt-out exception for AI training was dropped in the government's 2026 report after rights holders objected.
SoundExchange's 2024 announcement of a voluntary AI training-consent registry set a Q1 2025 target its own materials have not since confirmed as met.
Universal Music Group's 2023 filing sets out two principles: only humans hold copyright, and AI training on copyrighted works requires a negotiated license.
Suno's CEO argued in a 2024 blog post that training on copyrighted music resembles a musician learning a genre, not the infringement labels allege.
IFPI's Global Music Report 2026 attributes an eleventh straight year of growth to paid streaming while naming AI licensing and streaming fraud as watch items.
Goldman Sachs Research's 2025 Music in the Air report forecasts revenue nearly doubling by 2035 and estimates AI music at 0.1% of the royalty pool.
CISAC and PMP Strategy's 2024 study models a 2028 revenue-risk scenario, not an observed market outcome.
Deezer's newsroom states AI-generated uploads exceeded 50 percent of daily deliveries for the first time in June 2026.
Spotify's newsroom states it removed over 75 million spammy tracks in twelve months, a self-reported figure, not an audit.
DistroKid's help-centre article ties AI-track uploads to the same ownership and anti-impersonation rules as any release.
CD Baby's parent publishes a content policy naming unlicensed AI training data and artist mimicry as specific prohibitions.
Loudly's FAQ and licence agreement grant broad commercial use but forbid resale, sublicensing, and unauthorised film use.
Mubert's own pricing FAQ and licence page tie monetisation to the Pro and Business tiers and bar redistribution on any tier.
ElevenLabs' pricing and terms pages tie commercial rights to plan tier and reserve film, TV and radio for enterprise deals.
Stability AI's licence page sets a one-million-dollar annual-revenue threshold separating free and enterprise use of Stable Audio.
Suno's pricing page and September 2026 terms tie commercial exploitation to a per-tier monthly download allocation.
Udio's own help centre confirms downloads were disabled during its licensed transition with Universal Music Group, before any replacement shipped.
AIVA's own licence page ties copyright ownership, credit and monetisation rights to which subscription tier a composition was generated under.
Spawning's own tooling and its 2022 launch reporting show what the artist-led opt-out project does, apart from Herndon's separate Holly+ instrument.
Mycelia's own materials describe a prototype for embedding rights and payment data in music, years before AI made provenance urgent.
Bloom's App Store listing and its creators' own site show a tap-driven generative system, distinct from the 1996 Generative Music 1 disc.
Cave's own Red Hand Files letter states a ChatGPT song in his style is replication, not songwriting, and explains why in his own words.
The BBC's own interview quotes Sting warning of a fight to defend human creative capital against AI, days before his Ivor Novello fellowship.
NME's contemporaneous reporting quotes Guetta describing how he built an unauthorised AI Eminem-style verse and why he will not release it.
Kits AI's own ethics, earn and terms pages describe artist-submitted, revenue-sharing voice models rather than voices captured without consent.
Voice-Swap's own pages show a credit-based personal tier alongside a contract-backed enterprise path for licensed voices.
Moises' own help-centre articles state users keep output ownership while its generative models train on separately licensed data.
Endel's own press release for its Grimes collaboration credits her original vocals and music as the seed material its engine adapts.
ISMIR's own pages date its community to 2000, its incorporation to 2008, and list a downloadable paper archive for every year since.
ITU-R BS.1534 defines the hidden reference and anchors that make a MUSHRA listening test a specific, checkable method, not a label.
The original FAD paper defines a reference-free metric built for music enhancement, with a stated correlation to human perception.
The Hybrid Transformer Demucs paper reports its 9.20 dB SDR score on MUSDB HQ, a figure from the authors' own evaluation, not a third party's.
The mir_eval paper documents why inconsistent implementations of standard MIR metrics were confounding comparisons before the toolkit shipped.
Tzanetakis' own dataset page describes GTZAN's ad hoc origins, and a later peer-reviewed audit catalogued repetitions and mislabellings.
Two experiments found editing an AI poem lowered creativity, but co-creating with it restored creative self-efficacy.
An SSRN working paper found belief about a track's AI origin, not just its audio, changed how expressive listeners judged it.
A 152-listener preprint found human-composed clips rated more effective at eliciting a target emotion than AI ones.
A Frontiers study of 405 Chinese music majors found GAI-supported collaboration raised several self-reported creativity measures.
A 2026 paper reviews AI composition methods and proposes a new multi-level neural network framework, per its own abstract.
FakeMusicCaps pairs five text-to-music models' output with their MusicCaps prompts to benchmark detection classifiers.
C2PA's version 2.3 specification sets out exactly how content credentials embed into MP3, FLAC and WAV files.
DDEX's Electronic Release Notification standard structures how labels send release metadata to streaming services.
The ISRC Handbook shows the code identifies a recording but explicitly does not establish copyright or authorship.
ISWC registration requires uniquely identified creators and is explicitly unrelated to a work's copyright status.
DeepMind's Nature paper and open-source release cover text watermarking, distinct from its separate audio watermark for Lyria and NotebookLM.
ACRCloud's own documentation describes a fingerprint-matching and AI-detection service that feeds rights decisions made by others.
IRCAM's CUIDADO-era research supplied a descriptor taxonomy that MPEG-7 adopted, work that predates and underlies today's audio-recognition tools.
ITU-R BS.1770 defines the K-weighted loudness algorithm and LKFS unit that platforms build their own normalization targets on top of.
AcoustID's open Chromaprint algorithm and free lookup service identify already-known recordings by fingerprint, not by how audio was produced.
General MIDI standardized 128 instrument patches and drum mapping on top of the 1983 MIDI 1.0 protocol, unrelated to a file's authorship.
AES's own forensic-audio standards define authentication and evidence-handling methods for analog and digital recordings, not AI-detection.
Spawning's own tooling shows a voluntary opt-out signal that a cooperating trainer's pipeline can check, with DeviantArt a documented adopter.
Fairly Trained's published criteria define what licensed, public-domain or owned training data must look like to earn its certification mark.
ISNI assigns persistent IDs to named creators and entities; any human-versus-AI flag would sit in a separate metadata field, not the ID itself.
Boomy's own support articles say only Creator and Pro downloads carry full commercial rights, a policy that replaced exclusive distribution.
The MERT paper pairs an acoustic and a musical teacher model to pretrain a music-understanding network across 14 benchmark tasks.
Try another category or clear your search.
