Composing is not the same as producing sound
Making a computer produce sound and making it decide which notes to play are different problems. This article is about the second one. Early programs did not invent melodies out of nothing. Their authors wrote rules, probabilities or both, and the machine carried them out faster and more consistently than a person could.
The idea is older than electronics. In the eighteenth century, popular musical dice games let amateurs assemble a waltz by rolling dice to choose between prewritten measures. Every possible result was already built into the table of options, and chance picked the path. Ada Lovelace, writing about the Analytical Engine in the 1840s, also remarked that such a machine might in principle compose pieces of music if the rules of harmony could be expressed in its notation. It was speculation, since the machine was never completed, but it shows the idea was in the air long before there were computers to try it.
The Illiac Suite
The best-known early experiment came in the mid-1950s at the University of Illinois, where Lejaren Hiller, a chemist and composer, worked with Leonard Isaacson on a computer called the ILLIAC. Their output, a work for string quartet usually called the Illiac Suite, was completed around 1957.
The method was essentially generate and test. The program produced random musical material, then checked each candidate against rules and discarded what broke them. Early sections used strict rules similar to those taught in traditional counterpoint exercises. Later sections loosened the rules, brought in chromatic material, and experimented with other ways of controlling the output. Human performers played the finished score on ordinary instruments; the computer produced notation, not sound. Hiller and Isaacson also adjusted and arranged the results, so the work was a collaboration rather than a purely automatic product.
Probability as a compositional tool
A second major thread used probability more directly. The composer Iannis Xenakis, who had a background in architecture and engineering, developed what he called stochastic music in the 1950s, using ideas from statistics to shape large masses of sound. In the early 1960s he began writing programs, on an IBM computer available to him in Paris, to calculate the details of his compositions. The computer did not decide the style. Xenakis chose the distributions, ranges and overall form, and the program worked out the individual events, which he then used as material for scores.
Other composers used Markov chains, a mathematical way of choosing the next item based on a few preceding items. If a program learns that certain notes tend to follow certain other notes in a body of music, it can generate new sequences that resemble the source. The results usually sound locally plausible but wander without direction, because a chain with short memory has no sense of the overall form. That limit is part of why many composers have treated such programs as a source of raw material rather than finished pieces.
Chance without computers, and with them
Chance had also become an accepted compositional device for composers who never touched a computer. John Cage, for example, used chance procedures such as coin tosses in works from the 1950s. In the late 1960s Cage collaborated with Hiller on a large multimedia piece, HPSCHD, which used computer programs to carry out chance operations on material drawn from harpsichord music. It is often cited as an example of how experimental composers adopted computers as tools for managing complicated chance procedures.
Later developments
The generating and testing approach did not disappear. Through the 1970s and 1980s, researchers explored composition programs built on grammar-like rules, and a few built systems that analyzed existing composers’ music to produce new pieces in the same general style. Such programs can be impressive, but they depend on the music chosen as input and on human selection of the best outputs. A theme that runs through these projects, and through broader artificial intelligence research, is that capturing the rules of a style is easier than capturing the judgment that makes a piece worth hearing.
Does the result count as music?
The question has been debated since the 1950s, and reasonable people disagree. A few common positions are worth stating fairly.
- The skeptical view: a program only recombines patterns it was given, so any quality comes from its author or from the music it was fed.
- The tool view: composers have always used systems, from counterpoint rules to twelve-tone rows, and a computer is a more powerful way of working within one. The composer still decides what to keep.
- The listener view: music is defined by what people hear and value, not by how it was produced. If a piece moves an audience, how it was made matters less.
Part of the disagreement is about words. If “composer” means someone who makes the decisions that shape a piece, then the person who wrote the rules, chose the settings and picked the results has a strong claim. If it means someone who intends each note, early programs fit less well.
What this early work left behind
Early generative experiments showed that musical structure could be described precisely enough to be programmed, and just as clearly showed where such descriptions fell short. That tension remains useful. Many modern tools for arranging, accompaniment and sound design quietly use ideas that began as these experiments, with a human still steering the outcome.