MusicGen

MusicGen is Meta AI's single-stage auto-regressive Transformer model for high-fidelity, controllable music generation, driven by text prompts or melody inputs.

MusicGen, developed by Meta AI, is a powerful text-to-music generative model built on the AudioCraft library. It employs a single-stage auto-regressive Transformer (available up to a 3.3B parameter size) that operates over compressed discrete music tokens (EnCodec). This architecture efficiently generates high-quality audio, bypassing the need for cascading multiple models. The system excels at conditional generation: users provide a text description (e.g., 'lo-fi slow bpm electro chill') or an existing melody (chromagram) to guide the output, ensuring precise control over the final musical piece.

https://metademolab.com/audiocraft/musicgen/