AudioCraft

We’re introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artistic styles. We’re releasing the model weights and code, along with a tool to explore the generated samples. Provided with genre, artist, and lyrics as input, Jukebox outputs a new music sample produced from scratch. Jukebox produces a wide range of music and singing styles and generalizes to lyrics not seen during training. All the lyrics below have been co-written by a language model and OpenAI researchers. When conditioned on lyrics seen during training, Jukebox produces songs very different from the original songs it was trained on. We provide 12 seconds of audio to condition on and Jukebox completes the rest in a specified style. We chose to work on music because we want to continue to push the boundaries of generative models. Jukebox’s autoencoder model compresses audio to a discrete space, using a quantization-based approach called VQ-VAE.

AudioLM is a pure audio language model that generates high‑fidelity, long‑term coherent speech and piano music by learning from raw audio alone, without requiring any text transcripts or symbolic representations. It represents audio hierarchically using two types of discrete tokens, semantic tokens extracted from a self‑supervised model to capture phonetic or melodic structure and global context, and acoustic tokens from a neural codec to preserve speaker characteristics and fine waveform details, and chains three Transformer stages to predict first semantic tokens for high‑level structure, then coarse and finally fine acoustic tokens for detailed synthesis. The resulting pipeline allows AudioLM to condition on a few seconds of input audio and produce seamless continuations that retain voice identity, prosody, and recording conditions in speech or melody, harmony, and rhythm in music. Human evaluations show that synthetic continuations are nearly indistinguishable from real recordings.