Google Introduces MusicLM, a Generative Neural Network for Music Generation

Google has recently presented MusicLM, a generative neural network that is able to create music based on text and image input.

Google Introduces MusicLM

Currently, access to the network is not yet available to the public. The network was trained on 280,000 hours of music and is able to generate tracks in various genres, taking into account nuances in the input descriptions. For example, it can be asked to create a track that evokes a “feeling of being in space” or a “main soundtrack for an arcade game”.

MusicLM can also be based on existing melodies that the user sings, plays, or hums. Additionally, the system can interpret multiple sequential descriptions to create a longer track.

Furthermore, tasks for MusicLM can be set by combining an image and its caption, setting the level of experience of the virtual “musician”, or generating the sound of a specific instrument.

Finally, the system can create vocal parts, but they are often not full-fledged lyrics, but rather an approximation.