> ## Content Index > Fetch the complete content index at: https://onthehrzn.berklee.edu/llms.txt > Use this file to discover other available public pages before exploring further. # A History of Generative Music - URL: https://onthehrzn.berklee.edu/a-history-of-generative-music/ - Published: 2026-10-01T15:06:00.000Z - Updated: 2026-10-08T13:20:15.000Z - Description: From parlor-game dice rolls and analog tape loops to computer viruses and 24/7 infinite death metal streams, music’s dance with automation has been eccentric, hypnotic, and visionary. - Author: Adam Houghtaling - Tags: history of the future of music, generative music, generative ai, neural networks Generative music feels like such a contemporary topic, it’s hard to grasp that it dates back to the 11th century. But unlike the history of most musical forms, the story of generative music is one of discovery, obscurity, and rediscovery, with the actual term “generative music” not even appearing until the mid-90s. While ancient automata (see Ctesibius), Pythagorean tuning ratios, and the Chinese method for generating the twelve-*lü* pitch system all occurred way back in the BCE, long before the modern era, none meet the definition of generative music making. Even the earliest candidate for the generative music label, laid out in Guido of Arezzo’s 11th century music theory book, *Micrologus*, is a topic of debate among scholars. Also up for debate? The very definition of generative music. --- ## What Is Generative Music? Before 1965 there was no established umbrella term “generative music” to argue about. Centuries of dice games and melody hacks had their own labels: - Guido's disputed 11th-century vowel-based *modus* *inveniendi cantum* (The Method of Finding Chant) technique, which later scholars dubbed "mechanical composition." - Classical composer Josquin's 1503-1505 extension of Guido’s work was dubbed *soggetto cavato* (carved subject). - In the 18th-century, a popular generative music party trick famously performed by Mozart was called "the musical dice game." - Avant garde composer John Cage created generative works during the 1950s with a technique called "chance operations.” Then, in 1965, German philosopher Max Bense used the “generative” label to describe one-and-done algorithmic output initiated by a machine and the coinage stuck from an aesthetic point of view. In 1969, german academic Georg Nees used the term *Generative Computergraphik* to describe art created by executing algorithmic code. Well known for having coined the term “ambient”, which he defined on his 1978 album *Ambient 1: Music for Airports*, legendary musician and producer Brian Eno is less well known as coining and popularizing the label Generative Music in its modern musical sense in the mid-1990s. Eno described systems that could “produce original music forever,” lamenting the limitations which forced him to simply record the output of these systems as traditional albums rather than “sell the system itself, so that a listener would know that the music was always unique.” Artist and theorist Philip Galanter's widely sited 2003 definition is much broader than Eno's as it covers any art practice where the artist cedes control to a system operating with some autonomy, be it a computer program, a set of written instructions, or a chemical process. Throughout the relatively niche history of generative music inventors and engineers have sought to devise creative machines and ways for non-musicians to compose, facing scorn and heated opposition along the way. The hyper-scaling of generative AI takes these notions to a new extreme, but at heart, you'll find, the story remains much the same. --- ## GENERATIVE MUSIC BY ANY OTHER NAME... ### **What is Aleatoric Music?** Coined by acoustician Werner Meyer-Eppler in 1955, aleatoric music relies on controlled chance operations or real-time performer choices within a set framework to determine pitches, durations, or sequences. ### **What is Stochastic Music?** Coined by composer Iannis Xenakis in 1955, stochastic music applies statistical probability functions and mathematical distribution laws to calculate complex, cloud-like soundscapes. ### **What is Algorithmic Composition?** Emerging in computer music literature in the mid-1950s, algorithmic composition uses formal mathematical logic or computer code to generate musical scores and structural rules. ### **What is Process / Systems Music?** Associated with composers Steve Reich and Michael Nyman, this approach sets a rigid rule set or physical mechanism—like out-of-phase tape loops—into motion to let the piece unfold automatically. --- ### **11th Century** ### **Mechanical Composition** ### **Guido of Arezzo** Yes, the same monk who gave us *solfège* (better known as "do-re-mi") and one of the most important music theory books ever produced in *Micrologus*, also invented a method where a composer maps the vowels a-e-i-o-u onto pitches of a scale, then generates a melody by applying those pitches to the vowels in a text. A composer maps the vowels onto the notes of a scale, looks up a passage of scripture, and applies pitches to individual syllables in the passage. Whether this is the first genuine generative rule—input a text, run it through a fixed mapping, output a melody deliberately "composed" by no one—or not is unclear with modern scholars interpreting the instructions as simply a chapter about training students to *appreciate* the way vowel sounds and pitch already correspond in existing chant. The field itself isn't settled on whether this is the first known instance of generative music or simply music appreciation pedagogy that looks algorithmic from a distance. It would be another 500 years before the world sees the next recorded instance of generative music. --- ### **1503** ### **Josquin des Prez** ### ***Missa Hercules Dux Ferrariae*** ### (Hercules, Duke of Ferrara Mass) A Franco-Flemish composer working with both Italian and French courts, Josquin is credited with originating the technique later named *soggetto cavato* ("carved-out subject") by the theorist Gioseffo Zarlino in his 1558 treatise *Le Istitutioni Harmoniche (The Harmonic Institutions)*. The method takes a word or phrase, extracts all vowels, and then maps each vowel onto a *solmization* syllable from our friend Guido of Arezzo's hexachord system (ut-re-mi-fa-sol-la, the ancestor of do-re-mi). Each syllable corresponds to a fixed scale degree so a melody is mechanically "carved out" of the text rather than freely composed. His best-known application is the *Missa Hercules Dux Ferrariae* (c. 1503), where the vowels in "Hercules Dux Ferrariae," the name and title of his patron, the Duke of Ferrara, generate the mass's recurring melodic subject. A particularly clear early example of rule-based melodic derivation comes from classical composer Josquin, who set a fixed rule in motion and relinquished compositional control to the system. It earns him a place as arguably the first named composer to use a generative technique deliberately and repeatedly rather than as a one-off curiosity, roughly 289 years before Mozart's dice game, which is often erroneously pegged as the beginning of generative music. --- ### **1650** ### **Athanasius Kircher** ### **Arca Musarithmic** A Jesuit polymath in Rome, Kircher described in Book VIII of his *Musurgia Universalis* a computation device that could automatically generate musical settings for four-voice chorus of a given text and it could generate music in several styles, from a simple homophonic texture like a hymn to a complex fugal texture, all by combining preset permutations of musical patterns. Physically, it was a box of wooden slats or rods, each printed with numerical tables representing musical fragments ("musarithms"); you'd select rods according to the meter and syllable-count of your text, and combining them per Kircher's rules produced correct, if generic, polyphony. Echoes of Suno here in the 17th century as the purpose of the machine was to enable non-musicians to compose church music. Multiple physical copies were built and survive today. This predates Mozart's dice game by 130+ years and is arguably the first true generative-music machine in the sense of a rule-governed device (rather than a human) doing the compositional work, even as a human operator was still necessary to run the device. --- ## 1787 ### Wolfgang Amadeus Mozart ### *Musikalisches Würfelspiel* ![](https://storage.ghost.io/c/9a/8a/9a8a7ecd-106b-49ad-9d29-f0e9e57182b3/content/images/2026/09/musical-dice-game-1.jpg) Attributed to Wolfgang Amadeus Mozart in 1792, the *Musikalisches Würfelspiel* (Musical Dice Game) allowed anyone to "compose" a custom 16-bar waltz by rolling dice to select pre-written musical fragments from a reference grid. Because each fragment shared the same underlying harmonic structure, any random sequence produced a stylistically cohesive piece, generating over 700 trillion possible variations. In doing so, the game anticipated aleatoric and generative music centuries before John Cage used coin flips for chance operations or Brian Eno conceptualized self-generating soundscapes. When Mozart rolled the bones it transformed the role of the composer from authoring a single static score to engineering a rule-based system—an engine where human constraints ensured structural elegance while random chance drove the execution. This 18th-century parlor game shares its core mechanics with modern generative AI. When a neural network or large language model samples the next token, it is essentially rolling dice based on calculated probabilities. The game's grid of 176 fragments served as a loose structural analogy: constrained choice among alternatives introduces controlled variation. Today’s AI models operate on a vastly scaled-up version of a similar principle, using high-dimensional data distributions instead of printed tables, while temperature and probability filters fine-tune the randomness. Whether throwing dice on a tabletop or processing probabilistic calculations on GPUs, generative music has always relied on the same fundamental paradigm: a human designer creating the framework, and chance driving the output. --- ### **1821** ### **Dietrich Nikolaus Winkel** ### **The Componium** ![](https://storage.ghost.io/c/9a/8a/9a8a7ecd-106b-49ad-9d29-f0e9e57182b3/content/images/2026/09/winkel-compodium-generative-music.png) ****Dietrich Nikolaus Winkel** and what is arguably the first generative music machine, the Componium Built by Dutch/German craftsman Dietrich Nikolaus Winkel, [the Componium](https://www.mim.be/en/collection-piece/componium-0?ref=onthehrzn.berklee.edu) was an act of both genius and, reportedly, revenge. The story goes that years earlier, in 1814, Winkel had invented the mechanism for the first successful double-weighted musical metronome, a device the modest inventor failed to patent. Shrewd showman of mechanical wonders Johann Nepomuk Mälzel added a scale to Winkel’s invention, patented it as the "Mälzel Metronome," and took both the credit and the profit. In response, Winkel built the Componium in 1821—a large, self-playing mechanical instrument designed to imitate the sound of an entire orchestra and play endless variations on a theme—as a way to outclass Mälzel’s own feats as a mechanical engineer. Mechanically, the Componium consisted of two turning barrels taking turns performing two measures of randomly chosen music while the other silently slid horizontally to select the next variation. Contemporary audiences found it truly uncanny and suspected a human was hidden inside the instrument—a belief later assuaged by a panel of French scientists. Previous generative music milestones consisted of instructions or systems, but the Componium is arguably the first true generative-music making machine: a device that autonomously executes randomized selections and produces sound. It is still viewable at the Musical Instruments Museum in Brussels. Following his death an obituary noted that as “proof of his talents, we need only mention the Componium, or *Musical* *Improvisatore*.” --- ### **1913** ### **Marcel Duchamp** ### ***Erratum Musical*** ![](https://storage.ghost.io/c/9a/8a/9a8a7ecd-106b-49ad-9d29-f0e9e57182b3/content/images/2026/09/marcel-duchamp-generative-music.png) Marcel Duchamp; Photo: Public Domain Provocative Dadaist artist Marcel Duchamp—best known for a urinal called “Fountain” that has been called the most influential artwork of the 20th century—dabbled in chance operations to compose his *Erratum Musical* for three voices in 1912-1913\. The artist made three sets of 25 cards, one each for his sisters Yvonne and Magdeleine and for himself. Each card had a single note, the cards were mixed in a hat, and then drawn one at a time, composing by chance procedure and becoming a major influence on later chance operation maestro John Cage. The resulting piece was to be played without traditional emotional modulation, his goal a *"uniformity de rhythm, an accentuation."* Duchamp made a second, more elaborate piece the same year using numbered balls dropped into the moving wagons of a toy train as the chance mechanism. His attitude was telling: "Execution is rather pointless, in any event." For Duchamp, the generative system itself was the artwork, the sound was just its output. --- ### **1951** ### **John Cage** ### ***Music of Changes*** John Cage's arrival at generative music began not with philosophy but with a gift. In early 1951, the composer Christian Wolff gave Cage a copy of the *I Ching*, the ancient Chinese text used for millennia as a divinatory oracle built on 64 hexagrams. Cage recognized at once that its hexagram-finding chart was a better chance mechanism than the magic-square charts he had been using in works like the *Concerto for Prepared Piano*. He began writing *Music of Changes* immediately, a solo piano work in which nearly every compositional parameter—pitch, duration, dynamics, tempo—was surrendered to the *I Ching.* While *Music of Changes* remained unfinished, the shorter [*Imaginary Landscape No. 4*](https://www.youtube.com/watch?v=oPfwrFl1FHM&ref=onthehrzn.berklee.edu), scored for twelve radios, reached the concert stage first, premiering May 10, 1951\. *Music of Changes* was subsequently completed in four books by the end of 1951\. In 1952 Cage begin sitting in on D.T. Suzuki's now-famous Zen Buddhism seminars at Columbia University, which drew a remarkable cross-section of the New York intelligentsia. Suzuki's teaching did not originate Cage's method so much as furnish it with a philosophical vocabulary: the Buddhist case against ego and attachment gave Cage a framework for understanding, and advocating for, what he had already begun doing by instinct. He would later describe the discipline this demanded by saying, "I have developed a complicated composing means using I Ching chance operations, making my responsibility that of asking questions instead of making choices" (*An Autobiographical Statement*, 1990). --- ### **1955** ### **Iannis Xenakis** ### **Stochastic Music** Architect-turned-composer Iannis Xenakis treated sound not in terms of melody, but more like swarms of bees, ditching traditional harmony for high-level probability math. Xenakis developed stochastic approaches to composition in the 1950s, borrowing the mathematical term *stochastic* (from the Greek *stochastikos*, meaning "aiming at a target" or "good at guessing") and applying it to music theory in his landmark essay, *"La crise de la musique sérielle"* ("The Crisis of Serial Music"). He defined "stochastic music" as sound architecture governed by mathematical laws of probability, statistical mechanics, and set theory rather than traditional harmonic rules or rigid serialism, which he later codified in his landmark 1963 treatise, *Musiques formelles*, published as a special edition of *La Revue musicale*. Rather than writing notes one by one, Xenakis used statistical laws to calculate thousands of individual musical events mapped out like mass sound clouds on graph paper before translating them into orchestral scores. "With the aid of computers, the composer becomes a sort of pilot," he wrote, "he presses buttons, coordinates command mechanisms, and controls the sound result." --- ### **1956** ### **Lejaren Hiller & Leonard Isaacson** ### ***Illiac Suite*** In 1956, two chemist composers, Lejaren Hiller and Leonard Isaacson, fed punched paper tape into a five-ton University of Illinois mainframe (the ILLIAC I) and was among the first to create computer-generated music. The result was the *Illiac Suite* *for String Quartet*, where the computer generated pitch sequences used instances of Markov chains and Monte Carlo chance methods—a mathematical technique that uses repeated random sampling to estimate the probability of different outcomes—and screened them against strict rule-sets of traditional counterpoint. The piece proved that a cold, vacuum-tube monster could absorb centuries of human music theory and generate new musical compositions without a single human finger touching an instrument. Its first public performance brought audience reactions we may find familiar in their concern that “it does away with the need for human composers”. --- ### **1964** ### **Gottfried Michael Koenig** ### **Project 1 (PR1)** Avant-garde pioneer Gottfried Michael Koenig wanted to eliminate personal intuition from composition entirely, turning instead to systematic computer logic and transforming the composer from a tormented artist waiting for divine inspiration into a "composer of rules." Working at what would later be called the Institute of Sonology in Utrecht during the 1960s, Koenig wrote [*PR1* (Project 1)](https://koenigproject.nl/project-1/?ref=onthehrzn.berklee.edu) and [*PR2* (Project 2)](https://koenigproject.nl/project-2/?ref=onthehrzn.berklee.edu), pioneering computer programs that generated musical scores by applying mathematical rules to parameters like pitch, duration, and dynamic entry times. [He laid out his philosophy](https://www.thedkprojection.com/downloads/thoughts/Interview%5FKoenig%5FII.pdf?ref=onthehrzn.berklee.edu#:~:text=Indeed%2C%20I%20always%20have%20seen%20Project%201,other%20because%20of%20the%20built%2D%E2%80%90in%20random%20generator.) bluntly in 2014: "Indeed, I always have seen Project 1 as a composition. It is a computer program, but also a composition, only one that does not appear before endless variations of it." --- ### **1965** ### Steve Reich ### *It's Gonna Rain* Minimalist composer Steve Reich stumbled into a groundbreaking musical process in 1965 with little more than a couple of cheap tape recorders and a street preacher's Great Flood warning. After recording the shouts of Pentecostal preacher Brother Walter in San Francisco's Union Square in 1964, Reich isolated the spoken phrase "It's gonna rain" and began looping the brief fragment. He recorded the results and replayed both on two separate machines. As the motors naturally drifted out of sync Brother Walter’s end-times alarm began to overlap and morph into a dense, hypnotic and rhythmic cloud. Reich had accidentally stumbled on the phasing technique that would become a foundational system for continuous, evolving audio patterns. Reich captured his structural worldview in his famous essay [*Music as a Gradual Process*](https://ciufo.org/classes/114%5Ffl11/readings/reich%5Fprocess.pdf?ref=onthehrzn.berklee.edu): "I do not mean the process of composition, but rather pieces of music that are, quite literally, processes." --- ### **1965** ### **Alvin Lucier** ### ***Music for Solo Performer*** ### Perhaps the most surreal story in the history of generative music belongs to composer Alvin Lucier’s 1965 work, [*Music for Solo Performer*](http://www.alvin-lucier-film.com/solo%5Fperformer.html?ref=onthehrzn.berklee.edu). In the early sixties, Lucier was in a creative crisis, recovering from physical exhaustion and struggling with an involuntary speech stammer that made traditional performance excruciating. A chance meeting with an eccentric theoretical physicist and cyberneticist working for the Air Force Cambridge Research Laboratories led to experiments turning the composer’s alpha state brainwaves into sound and pushing them through raw speaker cones which then brushed against drum heads and percussive instruments. *Music for Solo Performer* premiered at Brandeis University's Rose Art Museum in 1965 with staging that looked less like a concert and more like an avant-garde séance. The performance was a test of meditative fortitude, as keeping the generative loop alive required Lucier to remain in an alpha brain state, indifferent to his own thoughts, the room, the audience, or the performance. It proved that a generative musical system didn't need algorithms or dice—it could simply turn the physical feedback loop of the human nervous system inside out. Lucier’s 1965 experiment is universally recognized as the first composition and performance to make use of brainwaves in history. --- ## **1969** ### **Raymond Scott** ### **The Electronium** Devo's Mark Mothersbaugh on Raymond Scott's Electronium Described as an "instantaneous composing machine," engineer and composer Raymond Scott was already a well-known space-age musical inventor when he began working on [the Electronium](https://www.raymondscott.net/em-article-2001/?ref=onthehrzn.berklee.edu) in 1959 as a way to eliminate human error from performance and make musical reproduction error-free. The machine he envisioned would allow a human to input a basic melody or prompt with the machine autonomously generating unpredictable, real-time compositional accompaniments in response via randomized sequences of tones, rhythms, and timbres. Completed a decade later in 1969, Scott described the the Electronium as a "Music Structure Generator" and said it was "capable of a seemingly inexhaustible palette of musical sounds and colors, rhythms, and harmonies." Berry Gordy, founder of Motown Records, was so taken by Scott and the potential of the device that he bought the only known machine in 1970 and then, in 1972, hired Scott as the label's Director of Electronic Music Research and Development to explore possibilities in automated hit-making. While Scott denied it was a prototype synthesizer as it had no keyboard, it was as one of the first machines to create music by means of artificial intelligence and its importance in pointing the way towards the electronic compositions of the future is undeniable. Devo co-founder Mark Mothersbaugh now owns the Motown Electronium, and has painstakingly brought it back to life. --- ### 1981 ### **David Cope** ### **Experiments in Musical Intelligence (EMI)** Facing composer's block while working on an opera commission in the early 1980s, University of California, Santa Cruz composer David Cope began building a program that could compose in his own style. The result was Experiments in Musical Intelligence (EMI), nicknamed "Emmy," which he developed over the following decades. EMI analyzed scores by a single composer, deconstructed them into short recurring fragments Cope called "signatures," and recombined them according to rules it extracted about how the composer put those fragments together. Cope called the method "**recombinancy**." Fed the works of Bach, Mozart, Chopin and others, EMI generated new pieces in their styles. Cope published the method in *Computers and Musical Style* (1991), and recordings of EMI's output followed in the 1990s. EMI also set off a fight. In a widely cited 1997 test at the University of Oregon, audiences were played a genuine Bach-style piece, a human imitation, and an EMI piece, and many judged the EMI work the real Bach. Following decades of backlash from musical scholars and critics who accused him of destroying human artistry, and others requesting he use EMI to complete notoriously unfinished works, Cope retired the original program around 2003, and purportedly destroyed the bulk of Emmy's databases soon after. Also in 2003, Cope built a successor to Emmy and named it [Emily Howell](https://www.youtube.com/@emilyhowell2359?ref=onthehrzn.berklee.edu), which was designed to compose in an original, contemporary-classical style that did not imitate existing composers. --- ### 1996 ### Brian Eno ### *Generative Music 1* In 1996, Brian Eno tried to do what he had complained about in earlier years: release the system instead of a recording of it. *Generative Music 1* shipped on floppy disk, playable through SSEYO's Koan software, which ran rules Eno had set up on the listener's own computer. The music never repeated, so each listener heard something different, and Eno said he had been drawn to the idea that the piece would keep changing indefinitely. It wasn't his first self-running system. *Discreet Music* (1975) used tape loops of different lengths, and *Music for Airports* (1978) used loop-based generative composition/production processes. But *Generative Music 1* was the first time he distributed the system, and it gave the term "generative music" its popular meaning. --- ### **2001** ### **François Pachet** ### ***The Continuator*** ### French computer scientist and musician [François Pachet](https://www.francoispachet.fr/?ref=onthehrzn.berklee.edu) wanted a machine that could jam with a human musician in real time, matching their unique style without feeling pre-programmed. Built at the Sony Computer Science Laboratory in the early 2000s, *The Continuator* used Markov models—a mathematical system that uses probabilities to predict how a condition changes from one moment to the next—to listen to a live performer on a MIDI keyboard, instantly learn their phrasing, and seamlessly take over the improvisation whenever the human paused. Pachet turned the computer from a static playback tool into an active, conversational collaborator that blurred the line between human expression and machine response. Jazz saxophonist Bernard Lubat tested it and marveled at the eerie interactive loop: "It is both me and not me... It knows me better than I know myself." --- ### **2006** ### **Joseph Nechvatal** ### ***Viral symphOny*** Beginning in the late 1990s, media artist Joseph Nechvatal began collaborating with computer scientists to unleash custom computer virus software onto digital sound files of human intimacy, classical compositions, and ambient noise, allowing the code to attack and degrade the audio in real time. Nechvatal described his self-contained artificial-life/virus-inspired software as "a collaborative electronic noise music symphony." The immersive and chaotic result was *Viral symphOny*, which redefined music not as a carefully constructed architecture, but as a living organism fighting a fatal infection. "I never involved myself in the actual hacking of computer systems," Nechvatal told [ASAP Review](https://asapjournal.com/feature/digital-viral-magical-a-conversation-with-joseph-nechvatal-joseph-nechvatal-and-jonathan-p-eburne/?utm%5Fsource=gemini#:~:text=There%20were%20all%20kinds%20of%20viral%20metaphorical,me%20it's%20an%20artistic%20form%20of%20expression.) in 2023\. "I never wanted to lose track of the virus as an artistic and poetic tool. My art virus has always been self-contained—within my own computers. For me it's an artistic form of expression." --- ### 2016 ### Pierre Barreau ### **AIVA (Artificial Intelligence Virtual Artist)** ### Inspired by a scene in the 2013 Spike Jonze film, *Her*, Luxembourg-based startup founder Pierre Barreau launched AIVA (Artificial Intelligence Virtual Artist) in 2016, feeding centuries of classical scores by Bach, Beethoven, and Mozart into deep neural networks. Instead of merely spitting out fragments, AIVA analyzed mathematical patterns in notation to compose full symphonic works so convincingly that in late 2016 France’s SACEM (Society of Authors, Composers and Publishers of Music) registered AIVA as the first virtual entity officially recognized as a composer. Moving generative music from academic experiments onto Hollywood scoring stages, the AI composed pieces performed by real orchestras for film trailers and commercials. During his 2018 TED Talk, ["How AI Could Compose a Personalized Soundtrack to Your Life,"](https://www.ted.com/talks/pierre%5Fbarreau%5Fhow%5Fai%5Fcould%5Fcompose%5Fa%5Fpersonalized%5Fsoundtrack%5Fto%5Fyour%5Flife?ref=onthehrzn.berklee.edu) Barreau explained how he saw the future of personalized music, and wondered with the strange sense of imagination and hubris the stock and trade of the tech world, what it may have been like "if someone like Martin Luther King, for example, had a personalized AI composer. Maybe then we would remember 'I Have a Dream,' not only as a great speech, but also a great piece of music, part of our history, and capturing Dr. King's ideals." --- ### **2016** ### **Google Magenta Project** In 2016, Google Brain researcher Douglas Eck launched the [Magenta Project](https://magenta.withgoogle.com/?ref=onthehrzn.berklee.edu) with a straightforward, provocative premise: Could deep learning models move beyond sorting data and actually make art? Built on TensorFlow as an open-source research initiative, Magenta spawned groundbreaking generative tools from [NSynth](https://magenta.withgoogle.com/nsynth?ref=onthehrzn.berklee.edu), which forged brand-new instrument textures through neural audio synthesis, to [Music Transformer](https://magenta.withgoogle.com/music-transformer?ref=onthehrzn.berklee.edu), which mastered long-range musical structure. Rather than locking its code behind corporate doors, Google released the software to the public, fostering an ecosystem of interactive plugins for digital audio workstations. As Eck framed the project's ambition: The question Magenta seeks to answer is: Can machines create compelling music and art? If so, how? --- ### **2016** ### **Sony CSL** ### **"Daddy's Car"** In 2016, Sony Computer Science Laboratories in Paris dropped ["Daddy's Car,"](https://www.francoispachet.fr/daddys-car/?ref=onthehrzn.berklee.edu) a cheerful, psychedelic pop song composed by an artificial intelligence system called FlowMachines. Under the leadership of computer scientist François Pachet (see the Continuator above), researchers fed the system a database of lead sheets spanning global music history, then instructed it to generate a brand-new tune specifically in the style of The Beatles. The AI generated the harmony and melodic blueprint, leaving French producer Benoît Carré to arrange the track and write the lyrics. Pachet viewed the AI generator not as a threat to human songwriters, but as the ultimate creative foil: "The way I see it is it's more like assisting the Lennon that needs a McCartney". --- ### **2018** ### **Endel** Founded in 2018 by Oleg Stavitsky and a collective of artists and programmers, [Endel](https://endel.io/?ref=onthehrzn.berklee.edu) turned generative music into a real-time utility for personal wellness. Rather than serving up pre-recorded tracks, Endel’s generative engine creates custom, non-repeating ambient soundscapes driven live by the user's immediate environment—pulling in heart rate, weather, time of day, and location data to adjust tempo and frequency. The technology proved so commercially potent that Endel partnered with WMG's Arts Music division in a deal to release 20 algorithmically derived albums in 2019, making it essentially the first algorithm signed to a major record label. Stavitsky described the cultural shift behind their reactive audio: "Because the world is so crazy right now, people are now almost self-medicating with sound". --- ### 2019 ### **Dadabots** ### ***Relentless Doppelganger*** Musician-engineers CJ Carr and Zack Zukowski threw traditional MIDI notes out the window in 2017, choosing instead to point deep neural networks directly at raw audio files. Using a modified SampleRNN architecture, their project [Dadabots](https://dadabots.com/?ref=onthehrzn.berklee.edu) fed relentless albums of technical death metal—like [Archspire](https://metal-temple.com/news/archspire-music-used-as-template-for-ai-generated-death-metal/?ref=onthehrzn.berklee.edu) and [Dillinger Escape Plan](https://loudwire.com/the-dillinger-escape-plan-krallice-artificial-intelligence-albums/?ref=onthehrzn.berklee.edu)—into the network, forcing the AI to predict incoming audio waveforms sample-by-sample in real time. The resulting chaos spilled into [*Relentless Doppelganger*](https://www.youtube.com/watch?v=JF2p0Hlg%5F5U&ref=onthehrzn.berklee.edu), an infinite, non-stop 24/7 stream broadcast live on YouTube without human curation or editing. As CJ Carr noted regarding why hyper-fast technical metal worked so well for raw neural generation: "Our theory is the faster the blast beats, the more stable the net... You're hearing everything it makes". --- ![](https://storage.ghost.io/c/9a/8a/9a8a7ecd-106b-49ad-9d29-f0e9e57182b3/content/images/2026/10/musnet-2.png) ### **2019** ### **OpenAI** ### ***MuseNet*** A deep neural network capable of generating 4-minute, 10-instrument compositions in genres ranging from Country to Mozart. OpenAI introduced MuseNet in a 2019 post by researcher Christine Payne. A deep neural network trained to understand the structural language of music, MuseNet was built on a transformer architecture similar to GPT-2\. It was capable of creating four-minute compositions trained by next-token prediction over hundreds of thousands of MIDI files, effortlessly blending up to 10 different instruments across disparate genres from Chopin to Queen. The system demonstrated an uncanny grasp of long-term musical context, remembering thematic motifs across hundreds of measures. As Payne reflected on the model’s knack for creative synthesis: "It turns out that transformer models, which are great at language, are also remarkably good at understanding the grammar of music."