The rise of artificial intelligence (AI) offers exciting opportunities for the music industry. Today, tools capable of automatically generating musical compositions or instrumental tracks are emerging, but most of these technologies are aimed primarily at professional musicians and producers. However, a new innovation developed by LG AI Research is a game changer.

An innovative system for everyone

LG AI Research researchers have developed an interactive system that makes it easy to transform any user’s musical ideas into compositions. This system, as indicated by Sangjun Han, Jiwon Ham and their colleagues in their publication on the arXiv server, is based on a autoregressive transformer trained on large musical datasets, while providing an intuitive interface.

System operation

The principle of this system is based on the generation of symbolic music which focuses on short musical motifs. It generates four bars of multi-track MIDI music from music metadata. Using two major datasets, namely the Lakh MIDI Dataset and the MetaMIDI Datasetthe model has been nourished by more than 400,000 MIDI files.

To train this model, the team converted each MIDI file into a musical event representation format, known as REMI. This format has the advantage of encoding MIDI data into tokens that illustrate various musical aspects, such as pitch and velocity. Here are the key advantages of the REMI format:

• Ease of learning : Capturing music dynamics effectively for AI.
• Flexibility : Allows you to modify various musical aspects during training.
• Control : Offer users better control over the generated composition.

An interface accessible to all

In parallel with the creation of their model, the researchers designed a simple and intuitive user interface. This interface consists of two main elements: a sidebar and a central interactive panel.

Interface Features

In the sidebar, users can define the music elements they want to generate, including:

• Instruments to use
• The tempo of the song

After generating a track, it is possible to edit it directly in the central panel, whether that’s adding or removing instruments, or adjusting when musical elements should start playing . This approach aims to maximize creativity while ensuring easy access for users of all levels.

Validity proven by experience

The researchers evaluated the effectiveness of their model through experiments measuring model capacity, musical fidelity, diversity, and control. Beyond validation, they expanded the model and compared it with other music generators. The results indicate a notable superiority in terms of control and musical quality.

The system demonstrated its ability to reliably generate four-measure musical sequences, adequately meeting user specifications. Looking ahead, the team plans to further improve this system by:

• Extending the duration of generated music tracks
• Expanding specification options for users
• Optimizing the user interface for greater accessibility

In conclusion

With the evolution of AI, musical creation is becoming more and more accessible to everyone. This innovative system from LG AI Research does not require musical expertise to allow users to express their ideas, paving the way for new forms of musical creativity.

“Our model, trained to generate 4 measures of music with global control, has limitations in terms of extending music duration and controlling local elements at the measure level,” the researchers wrote. “However, our attempts are important to generate high-quality musical themes that can be used as a loop. »

More information: Sangjun Han et al, Flexible control in symbolic music generation via music metadata, arXiv (2024). DOI: 10.48550/arxiv.2409.07467

Journal information: arXiv