Meta Adds JASCO and MusicGen-Style to AudioCraft

Meta's Fundamental AI Research team has updated its AudioCraft repository, integrating two new specialized models named JASCO and MusicGen-Style. The framework brings together an array of tools for music creation, sound generation, neural audio compression, diffusion decoding, and watermarking into a unified PyTorch stack. By establishing shared configuration, training, and inference infrastructure, the toolkit aims to help practitioners evaluate, adapt, and build conditional audio architectures with less friction.

The JASCO model introduces flow-matching text-to-music generation guided by specific temporal conditions. Offered in 400M and 1B parameter sizes, it was trained on approximately 16,000 hours of licensed music. Practitioners can pass JASCO a global text description alongside precise, time-aligned local inputs such as chord changes, drum tracks, and structural melodies. Meanwhile, MusicGen-Style relies on a 1.5B parameter transformer conditioned on text prompts and a short audio sample lasting between 1.5 and 4.5 seconds to match reference styles. Running inference on MusicGen-Style requires a GPU with at least 16 GB of memory.

These new additions expand an ecosystem that already bundles several underlying components. The repository includes MusicGen's autoregressive transformer, which comes in 300M, 1.5B, and 3.3B parameter sizes, alongside AudioGen for sound effects, EnCodec for neural compression, MAGNeT for non-autoregressive masked generation, MBD for diffusion decoding, and AudioSeal for localized watermarking. Running the stack requires Python 3.9 and PyTorch 2.1.0. Meta released the underlying code under an open MIT license, whereas the official model weights are made available under a non-commercial CC-BY-NC 4.0 license.