Jukebox scraped a million songs and said so plainly

The system

Jukebox is a raw-audio music generation model that OpenAI introduced in a technical paper and a companion research post, both dated 30 April 2020. Instead of generating a symbolic score, it compresses waveform audio into discrete codes with a multi-scale VQ-VAE and predicts those codes with autoregressive Transformers, so the output includes vocals rather than only instrumental parts. A listener can steer generation with an artist and genre label, and can pass in unaligned lyrics the model attempts to sing. OpenAI published model weights, code and several thousand uncurated samples alongside the writeup.

What the documents establish

The paper states the VQ-VAE compresses 44.1kHz audio through three bottleneck levels, at 8x, 32x and 128x, each using a 2048-entry codebook, with the top-level prior further conditioned on artist, genre and lyric tokens. Both documents describe the same training corpus in near-identical language: OpenAI crawled the web to assemble 1.2 million songs, 600,000 of them in English, paired with lyrics and metadata — artist, album, genre, year and mood tags — taken from the fan-run site LyricWiki. Neither document calls that corpus licensed; it is described only as scraped. The paper also discloses performance limits the blog repeats more plainly: generating one minute of top-level tokens takes roughly an hour, and upsampling that minute to full-resolution audio takes about eight hours, while the top compression level loses high-frequency detail and gains artifacts comparable to a low-bitrate codec. The blog adds one fact the paper omits — ten musicians reviewed the model before publication.

Craft and rights

For a producer, the headline capability is real but narrow: Jukebox can render a rough vocal-and-instrumental sketch in a chosen style, but the hours-long render time for one usable minute rules out anything resembling an interactive session, a constraint both documents state outright. The rights question sits in what the training description leaves unsaid. A crawled, unlicensed corpus of 1.2 million songs paired with lyrics scraped from a fan wiki carries no documented consent or licensing basis in either source; OpenAI states what it collected, not what rights it held to collect it. That gap is this entry's editorial read, not a claim either document makes about infringement. The musician feedback the blog reports — that Jukebox's limitations kept it from fitting their creative process — is itself a modest, source-backed check against treating the release as production-ready.

Outcomes and open questions

Neither document revisits the training-data description after publication, so whether OpenAI later relicensed or restricted any of the material is not established here. The contrast with later disclosures is instructive: Meta's 2023 MusicGen paper names specific licensed catalogues, a level of detail Jukebox's authors did not publish.

• Does the tool's own documentation name the training sources, or only describe how they were gathered?
• Is a stated generation time compatible with the workflow you actually intend to use it for?
• When a lyric or vocal style is reproduced closely, whose consent was documented, if any?

Jukebox is a dated marker of what raw-audio generation could do in April 2020, and of how little training-data disclosure the field then considered standard.