OpenAI announces Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles
complete with lyrics Jordan Pearson / VICE : This Software Can Make Kanye West Rap Eminem's ‘Lose Yourself’ Raf Noboa y Rivera / The Miscellanies : Hope will give us power Tweets: @openai : Introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles. We're releasing a tool for everyone to explore the generated samples, as well as the model and code: https://openai.com/... https://twitter.com/... Andy Baio / @waxpancake : Open AI just announced Jukebox, a neural net that generates music *with singing* using genres, artists, and lyrics as input. The results are INCREDIBLE. https://openai.com/... @earacherecords : We've heard worse demos to be fair.. 😎 https://twitter.com/... Tim Soret / @timsoret : The replacement of human intellect & creativity is surely arriving, brick after brick. https://twitter.com/... @metaleptic : trying to figure out which AI-generated soundtrack goes best with the images. https://openai.com/... @amydentata : Someday we will have a Spotify that delivers 100% novel audio every second of every stream and we will all lose our minds from gibberish overload https://twitter.com/... @dougscripts : If you haven't checked this out, you really need to. It's quite amazing https://twitter.com/... Albert Wenger / @albertwenger : Another big step forward towards machine creativity https://continuations.com/... https://twitter.com/... @mark_riedl : Training neural networks on music with lyrics results in voice-like babbling. But condition that audio with lyrics and you can get a neural network to sing https://openai.com/... @mikerugnetta : I have claimed in the past that the uncanny valley exists only in theory, not practice. I am not so sure anymore. https://twitter.com/... Paolo Pedercini / @molleindustria : Wow, the researchers at Open AI are generating impressive songs with lyrics and by genre, starting from raw recordings: https://openai.com/... I'm very fond of this autocompleted Never Gonna Give You Up (a kind of sophisticated rickroll): https://soundcloud.com/... Jacob Ward / @byjacobward : In just the year and a half I've been at @NBCNews, AI has green-lit scripts, made auctionable art, and completed unfinished symphonies. Now it can do the whole songwriting and performing process. https://twitter.com/... Kyle McDonald / @kcimc : omg @OpenAI has entered the AI-generated music ring. there are sounds in here i've never heard before. this is a big moment. https://openai.com/... https://soundcloud.com/... Kent Bye VoicesOfVR / @kentbye : 1 Generalizes music from Unseen Lyrics co-written by a language model + @openAI researchers 2 Trained on lyrics, Re-Renditions produces different song 3 Trained on 12 seconds of audio then Completed 4 Fun songs = duet + popular song lyrics + deep learning songs + children's music https://twitter.com/... Je-Meng / @jemenger : This is unbelievable, terrifying really. Listen to the examples down the thread. It is as if a few Sims teamed up with HAL 9000, provisioned serious resources on AWS, and started a band. https://twitter.com/... Jonathan Mann / @songadaymann : I don't want to tempt fate or anything, but please, if I die, could someone make sure that I keep writing a Song A Day by using something this? Please? https://twitter.com/... Tom Brown / @nottombrown : Listening to this singing neural net over the last few months has dramatically increased my appreciation for country music 🤠 https://twitter.com/... Oriol Vinyals / @oriolvinyalsml : Very happy to see such a cool application of VQ-VAEs by our friends @OpenAI! Congrats! https://twitter.com/... Brendan Dawes / @brendandawes : This is only the beginning. https://twitter.com/... Josiah / @jgulden : Don't lie, these are Bon Iver lyrics https://twitter.com/... Greg Brockman / @gdb : Just released Jukebox, a neural network that generates music as raw audio. All of the sound you hear, including the singing, is generated: https://twitter.com/... Matthew Panzarino / @panzer : I miss the AI Kanye https://twitter.com/...
Context & Ripple Effects
Jukebox was OpenAI's first foray into raw-audio generation: a neural net producing complete songs with rudimentary singing in named genres and artist styles, released alongside the model, code, and a sample explorer. At the time it read as a research artifact — impressive demos like a Kanye-style rendition of Eminem lyrics, but no product attached.
The coverage since shows what grew out of that seed: Suno and Udio turned text-prompt song generation with vocals into consumer products by 2024, ElevenLabs extended synthetic voice into lyrics and song samples, and by late 2025 sources reported OpenAI itself building music-generation tools from text and audio prompts. Jukebox is the origin point of a market its own creator left to startups — and is now re-entering.
First-order effects
- OpenAI hands researchers and hobbyists an open model and codebase for full-song audio generation, making vocal synthesis reproducible outside the lab overnight.
- Artists whose styles the model imitates — the demos explicitly name genres and artists — have no opt-in mechanism, setting up the consent-and-licensing question before any commercial product exists.
Second-order effects
- Four years on, the technique Jukebox demonstrated becomes the core of venture-backed products: Suno, Udio, and ElevenLabs compete on turning prompts into finished songs with vocals, forcing rights holders and labels to confront synthetic tracks at scale.
- OpenAI's reported 2025 return to music tools means the lab that published the original research now competes with the companies that commercialized it, while its Voice Engine work keeps it adjacent on the voice side.
Third-order effects
- If the pattern holds, raw-audio generation follows the path of text and image models: open research release, startup productization, then incumbent re-entry — with platforms like Spotify already building policy layers such as AI Persona badges and recommendation exclusions to manage synthetic supply.
- The unresolved piece from Jukebox's launch — training on and imitating identifiable artists without permission — hardens into the central legal and licensing battleground for the entire AI music category.
The trend: Generative music is completing the loop from OpenAI's 2020 research release through startup commercialization back to incumbent labs, dragging licensing and platform-policy questions along with it.