If ChatGPT can write articles and DALL-E can draw, who says AI can’t sing? Artificial intelligence has taught us how to write a resume correctly, compose an email, draw a cat in the style of Picasso, and solve complex math problems with the ease of a child. That’s all great. But OPENAI is now trying to go beyond Excel spreadsheets and pretty emojis: create a song that people actually want to hear more than once with OPENAI JUKEBOX. While no official release date has been set, the same question is circulating in tech and research circles: Is OpenAI ready to launch a new version of Jukebox? OpenAI Jukebox 2.0? This is an experimental AI model that years ago proved that it could compose music, sing lyrics, and even imitate Elton John, Nirvana, or anyone who has ever performed on MTV—all without a single human note. Well, at least without vocal cords.
OpenAI Jukebox: AI can do more than just repeat the key of C major
When OpenAI first launched Jukebox, critics were skeptical, and for good reason. A model that generates music from scratch (including vocals) sounds like a futuristic scenario that a music producer with plenty of time and few artists would come up with. But Jukebox works. By analyzing more than a million songs (including lyrics, genres, artists, and even release years), he created a whole new way to make music: not based on instruments, but on data.
Instead of notes, the model produces sounds. And it does it directly in a studio-quality audio format. What you hear is a song that never existed, but someone could easily mistake it for a lost demo by Radiohead from 1998. Technically, Jukebox first encodes the raw sound into a compressed symbolic language – an approximation of musical DNA. This encoded recording is then processed by three layers of neural networks, each of which contributes to the melody, rhythm, vocals, and sound texture in its own way. The final step is reverse decoding – the AI folds the sound waves back to create sounds that humans can hear but the AI can only “predict.”
In recent weeks, developer activity in OpenAI’s audio division has increased. Some commit logs have appeared on GitHub that mention model optimizations. Some researchers have hinted in podcasts (and especially in informal conversations on X/Twitter) that they are developing new architectures that allow for faster generation, better vocal representation, support for multiple languages, and – especially interestingly – better understanding of song structure.
Conclusion: Sometimes the best sound comes from the silence before the premiere
No official confirmation has been made yet. OpenAI is silent, but its silence is unusually loud. The atmosphere among researchers is like a Kanye West album – no one knows when it will be released, but everyone is already writing reviews. If Jukebox 2.0 is really what the rumors predict, we may soon hear songs that are no longer based on personal inspiration, but on the collective memory of millions of songs. A sound we have never heard before, but so familiar.