MusicGen

MusicGen stands as a cutting-edge solution for simple and controllable music generation. This single-stage auto-regressive Transformer model is distinctive in its training approach, utilizing a 32kHz EnCodec tokenizer with 4 codebooks sampled at 50 Hz. Key features include:

Key Features:

• Single-Stage Auto-Regressive Model: Generates all 4 codebooks in one pass for efficient music creation.
• No Requirement for Self-Supervised Semantic Representation: Unlike existing methods, MusicGen doesn’t require a self-supervised semantic representation.
• Parallel Codebook Prediction: Introduces a small delay between codebooks, allowing for parallel prediction and reducing auto-regressive steps to 50 per second of audio.

Training Data:

• 20K Hours of Licensed Music: Trained on a vast dataset, including 10K high-quality internal music tracks, ShutterStock, and Pond5 music data.

Use Cases:

• Efficient Music Generation: Generate music with only 50 auto-regressive steps per second of audio.
• Controllable and Predictable Output: Achieve control and predictability in music creation with parallel codebook prediction.
• Diverse Music Styles: Explore a wide range of music styles with the influence of 20K hours of licensed music data.