by Meta
MusicGen by Meta AI is an open-source music generation model that creates music from text descriptions or melody inputs. Part of Meta's AudioCraft family, it uses a two-stage architecture combining EnCodec neural audio compression with a transformer-based language model. Trained on 20,000 hours of licensed music, it produces coherent stereo audio at 50kHz.
MusicGen generates from descriptions, not lyrics. Don't use [Verse]/[Chorus] tags:
Upbeat jazz with piano and saxophone, walking bass line,
brushed drums, warm and lively, 130 bpm, reminiscent of
a smoky New York club
Keep descriptions under 500 characters with key musical attributes:
Ambient electronic, soft pads, gentle arpeggios,
reverb-heavy, dreamy and floating, 72 bpm
Upload a reference audio file to guide the harmonic and melodic structure while the text controls the style.
| Parameter | Description |
|---|---|
| duration | Audio length in seconds (up to 30s) |
| temperature | Randomness 0-1.5 (default 1.0) |
| top_k | Token sampling pool (default 250) |
| top_p | Nucleus sampling threshold |
| cfg_coeff | Classifier-free guidance (default 3.0) |
Quick tips from the community about what works with MusicGen right now.
Sign in to share a tip.
No tips yet. Add a tip for this model.