AI is learning to read the paper trail of computer music

Claudio Ambrosini’s Cadenza estesa e coda, a 1981 work for flute and electronics, has a life on record. A Stradivarius release credits flautist Daniele Ruggieri alongside Alvise Vidolin for sound direction and Daniele Torresan for computer realisation. Those credits make a useful place to begin: the electronic part has a production history, too.[2]

An August paper from the University of Padua reproduces another trace of the piece: a Music V code printout with faded characters and shaded horizontal bands. It belongs to the material researchers are teaching AI to read.[1]

Music V was among the programs used for batch sound synthesis at Padua in the 1970s. The centre’s history describes an IBM mainframe linked to an IBM S/7 for four-channel conversion from digital data to audio. For a batch-synthesis piece, the computer calculated sound from instructions before the listener could hear the result.[3]

That history also records a separate interactive system, ICMS, with facilities for real-time synthesis, editing, mixing and spatial distribution. The distinction matters when reading an archive. A program name can identify a particular way of working, including the controls a musician had available. Treating all old electronic music as the output of one generic synthesiser would erase those choices.[3]

Recovering a way of making sound

In their 2022 history of the centre, researchers explain why preserving electroacoustic work involves operating practices as well as recordings. A recording carries the audible result; notes, software and equipment records help explain its construction. They let a later musician or scholar ask how a sound was produced and which decisions belonged to the people making it.[4]

The Music V project links recognised code with surviving manuals. Its proposed interpretation stage retrieves relevant documentation for a language model, then turns the program’s relationships into explanations and diagrams. The manuals supply the language’s local grammar and meaning.[1]

The working OCR component has roughly 25,000 annotated characters behind it and reports a preliminary character error rate of 10–15%. The paper has no completed benchmark across scores. Grammar correction remains under development; explanation and diagram generation are prototypes awaiting systematic expert validation.[1]

A printed instruction, its documented meaning and the resulting explanation are therefore separate objects to preserve. Keeping them connected gives the next researcher a way into the computer part of the composition.[4]

Report an error or suggest a correction