Compare AudioLM vs. OpenAI Jukebox

AudioLM is an innovative audio language model designed to create high-quality, coherent speech and piano music by solely learning from raw audio data, eliminating the need for text transcripts or symbolic forms. It organizes audio in a hierarchical manner through two distinct types of discrete tokens: semantic tokens, which are derived from a self-supervised model to capture both phonetic and melodic structures along with broader context, and acoustic tokens, which come from a neural codec to maintain speaker characteristics and intricate waveform details. This model employs a series of three Transformer stages, initiating with the prediction of semantic tokens to establish the overarching structure, followed by the generation of coarse tokens, and culminating in the production of fine acoustic tokens for detailed audio synthesis. Consequently, AudioLM can take just a few seconds of input audio to generate seamless continuations that effectively preserve voice identity and prosody in speech, as well as melody, harmony, and rhythm in music. Remarkably, evaluations by humans indicate that the synthetic continuations produced are almost indistinguishable from actual recordings, demonstrating the technology's impressive authenticity and reliability. This advancement in audio generation underscores the potential for future applications in entertainment and communication, where realistic sound reproduction is paramount.

We are excited to unveil Jukebox, a cutting-edge neural network designed to create music, including basic vocalization, in diverse genres and artistic expressions as raw audio. Alongside the release of the model weights and code, we are offering a tool to help users explore the music samples generated by Jukebox. By inputting genre, artist, and lyrics, users can receive entirely new music pieces crafted from the ground up. Jukebox is capable of producing a vast array of musical and vocal styles, and it can also generalize to lyrics that were not part of the training dataset. The lyrics included here have been collaboratively crafted by researchers at OpenAI and a language model. When provided with lyrics from its training set, Jukebox generates songs that diverge significantly from the originals, showcasing its creative capabilities. Users can input a 12-second audio clip for Jukebox to build upon, with the final output reflecting a desired style. Our focus on music stems from a desire to advance the potential of generative models further. Utilizing a quantization-based approach called VQ-VAE, Jukebox’s autoencoder model effectively compresses audio into a discrete latent space, enabling innovative sound generation. As we continue to refine these technologies, we look forward to the creative possibilities that lie ahead.