MusicGen
MusicGen is a text-to-music generation model developed by Meta AI, capable of generating high-quality musical audio from text descriptions and optional melodic input.
Definition
MusicGen is a text-to-music generation model developed by Meta AI, capable of generating high-quality musical audio from text descriptions and optional melodic input.
Overview
MusicGen was released by Meta AI (part of the FAIR research group) in June 2023 as a single-stage autoregressive transformer model for controllable music generation, trained on a licensed dataset of music and text/metadata descriptions. Unlike some earlier music generation approaches that required multiple cascaded models or hierarchical generation stages, MusicGen operates over discrete audio tokens produced by EnCodec, Meta's neural audio codec, and generates these tokens directly in a single pass, which simplifies the architecture and training pipeline while still producing coherent, high-fidelity audio. MusicGen accepts a text prompt describing the desired music (genre, mood, instrumentation, tempo) and can optionally be conditioned on a reference melody, letting users guide the harmonic and rhythmic structure of the output while the model fills in instrumentation, style, and arrangement. This melody conditioning is a distinguishing feature compared to purely text-conditioned systems, giving musicians and producers more precise creative control over generated output. Meta released MusicGen's code and pretrained model weights openly, along with a range of model sizes, making it one of the more accessible text-to-music systems for researchers and developers to run and fine-tune themselves, in contrast to closed, API-only competitors. MusicGen sits within Meta's broader generative audio research alongside AudioCraft (the umbrella framework that includes MusicGen, AudioGen, and EnCodec) and is commonly used for prototyping soundtracks, background music generation, and academic research into controllable generative audio models.
Key Concepts
- Single-stage autoregressive transformer generating discrete audio tokens
- Built on EnCodec, Meta's neural audio codec, for tokenizing audio
- Text-to-music generation from natural language prompts
- Optional melody conditioning for guided musical structure
- Released with open weights and code across multiple model sizes
- Part of Meta's broader AudioCraft generative audio framework