MusicLM

Architecture

• parameters: Proprietary (undisclosed)
• context_length: Generates several minutes of audio (24kHz)
• training_data: Trained on a large proprietary music dataset
• inference: Hierarchical Transformer (text-to-audio)

Capabilities

• Generates high-fidelity music from text descriptions
• Produces multi-minute compositions with complex structures
• Can condition on melodies (e.g. humming) to create stylized music

Benchmarks

• Quality: Outperforms prior music models in audio quality and prompt alignment

Safety

• Trained on licensed music data
• potential copyright considerations for generated content.

Deployment

• regions: global
• hosting: Google Cloud
• integrations: Not deployed; results reported in research publication only.

Tags