Google: MusicLM produces music after entering text

Google: MusicLM produces music after entering text

San Francisco, 1/30/2023

Google researchers introduced a new AI system that generates music from text descriptions. Dubbed MusicLM, the new model outperformed other AI music generators in both audio quality and its ability to stick to prompts. For copyright reasons, Google will not release MusicLM yet.

MusicLM has been trained with over 280,000 hours of music, making it capable of producing tracks of different genres, instruments and concepts. Users enter a natural language description of what music they want, including the number of instruments, and the system outputs high-fidelity audio. Users can also sing, whistle, or hum to build on a melody. The model „produces music at 24 kHz that remains constant for several minutes,“ according to the researchers.

The system is based on Google’s Transformer neural network architecture, which has proven particularly useful in the AI areas of natural language processing and computer vision.

Google researchers also released an AI training dataset with 5,500 pieces of music to help other researchers working on automated song generation