Suno

Contribution to AI

Singing was the part that had not worked. OpenAI's Jukebox could produce vocals in 2020 but took hours of compute per minute of audio, and Google's MusicLM, described in early 2023, handled instrumental texture while any lyrics dissolved into syllables. What went up on the web in December 2023, and reached a much larger audience through a Microsoft Copilot plugin at the same moment, was a system that turned a sentence into a finished mix — verse, chorus, and a lead voice singing words the user chose — in about a minute, free. The earlier open release, Bark, had shown the approach: treat audio as a token stream a transformer can predict, speech and laughter and noise alike, rather than modelling music as notes on a stave. Version 3 in March 2024 pushed generations to four minutes; version 4, that November, sharpened diction and added reusable voice personas. The consequences were legal as much as technical. Universal, Sony and Warner sued through the RIAA in June 2024, and the August answer conceded that training drew on recordings found across the open internet while arguing fair use — which turned an abstract argument about generative models and copyright into a concrete question before a court. Warner settled in November 2025 with a licensing agreement, establishing the template by which an unlicensed music generator becomes a licensed one.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).