Meta Unveils AudioCraft: Generative AI Models for Music and Audio Creation

Artificial intelligence is quietly infiltrating the music industry, creating viral songs, resurrecting favorite singers' voices, and even earning Grammy‑like recognition; Meta has released new AI tools that make music generation easier.

On Tuesday, Meta announced AudioCraft, a generative AI suite capable of producing "high‑quality and realistic" music directly from text.

AudioCraft comprises three models: MusicGen, AudioGen, and EnCodec. MusicGen and AudioGen generate sound from text—MusicGen creates music while AudioGen produces specific audio and sound effects.

You can access MusicGen on HuggingFace and interact with its demo, prompting it with descriptions such as "an 80s driving pop song with heavy drums and synth pads."

EnCodec is a neural audio codec that compresses audio and reconstructs the input signal; the latest version reduces artifacts and generates higher‑quality music.

Meta also released a pretrained AudioGen model that can generate environmental sounds and effects, like dog barks or floor creaks.

All three open‑source model weights and code have been shared, enabling researchers and practitioners to build upon them.

Meta suggests AudioCraft could become a new standard instrument, much like the synthesizer.

Meta adds that with more control, MusicGen could evolve into a novel instrument, similar to the early days of synthesizers.

This is not the first generative AI model of its kind; Google launched MusicLM earlier this year, and recent research reveals Google is also using AI to reconstruct music from human brain activity.