Meta launched a neural network that creates music from text

The MusicGen neural network, developed by Meta Corporation, has been launched. This advanced system generates music based on textual descriptions provided. For instance, users can request the AI to create “an energetic 80s pop track with prominent drums and synthesisers in the background” or “jazz music suitable for an elevator ride.”

Initially, a demo version of MusicGen is available on Hugging Face hosting. Later on, the open-source code will be shared on GitHub, where pre-trained models are already accessible.

To train MusicGen, Meta Corporation utilised a dataset consisting of 10,000 licensed songs and 390,000 instrumental recordings sourced from media libraries such as ShutterStock and Pond5. This neural network can generate melodies by interpreting descriptions as well as by learning from existing songs.

It is worth noting that Google previously introduced a similar neural network called MusicLM. According to its developers, MusicLM can create genres like “melodic techno,” “relaxed jazz,” and even “industrial techno sounds with repetitive and hypnotic rhythms.”

However, experts believe that the Meta neural network outperforms MusicLM in terms of music generation capabilities.