Jukebox, an open-source neural network marvel, stands as a creative powerhouse, conjuring music and even rudimentary singing in a diverse array of genres and artist styles, all delivered as pristine raw audio. This tool generously shares its inner workings, comprising both the code and model weights, alongside an exploration tool to uncover the treasures of the generated samples.

Unlocking a world of sonic possibilities, Jukebox empowers users to steer the creative process by providing input on genre, artist, and lyrics, yielding fresh musical samples in response. It’s a master of versatility, generating a spectrum of musical and singing styles while also adapting to unfamiliar lyrics, not encountered during its training.

Remarkably, Jukebox can craft music that defies resemblance to its original training data when guided by lyrics from its training set. This musical sorcerer allows users to furnish a mere 12 seconds of audio, with the tool gracefully completing the rest in the user’s chosen style.

Jukebox takes an audacious approach by modeling music directly as raw audio, despite the considerable challenge posed by the sheer length of raw audio sequences. To overcome this hurdle, it employs an autoencoder to condense raw audio into a lower-dimensional realm, where it weaves its auditory magic before gracefully up-sampling back to the realm of raw audio.

Jukebox represents an adventurous frontier in the realm of generative models, offering greater expressive prowess than counterparts that generate music in symbolic piano roll notation. It beckons to those eager to embark on a sonic journey of AI-generated music experimentation.