Generate music and sound from text

Generate music and audio from text descriptions at multiple scales

Applications requiring controllable, variable-length music generation without licensing constraints.

• · Text descriptions (e.g., 'happy upbeat electronic dance music')
• · Optional melody conditioning (for MusicGen-melody)
• · Optional reference audio (for style transfer)

• · Stereo WAV audio files (32kHz for MusicGen, 16kHz for AudioGen)
• · Variable duration (1-30 seconds)

• · audiocraft library
• · PyTorch 2.0+
• · Hugging Face Transformers 4.30+

• · GPU required (CUDA/Metal)
• · 8GB+ VRAM for large models
• · Sufficient disk for model weights (3.3B max)

• · Memory overflow on small GPUs
• · Generated audio may not match description precisely
• · Melody conditioning requires careful audio preprocessing
• · Temperature/cfg_coef tuning needed per use case

• · Meta-built
• · Production model sizes from 300M to 3.3B parameters
• · Stereo support
• · Multiple conditioning modes documented