14th June 2023, Mumbai: Meta has introduced MusicGen, an artificial intelligence (AI) music-generating system that can create music based on text prompts or melodies. Similar to Google’s MusicLM, which builds on existing melodies, MusicGen has been trained on 20,000 hours of music to tackle the complex task of generating harmonies and melodies from different instruments. A demo of MusicGen was released on Hugging Face, showcasing its capabilities.
To test the capabilities of MusicGen, Interesting Engineering asked the system to generate a “rock lullaby.” After 341 seconds, the system produced a 15-second audio clip that combined tunes from the piano and an electric guitar. The result sounded reminiscent of the beginning of an old classic rock track.
Meta and its team of researchers have compared MusicGen with other AI music-generating models such as MusicLM, Riffusion, and Mousai. They uploaded a sample set and evaluated the models using the same textual prompts to analyze their performance.
In line with its open-source approach, Meta has made MusicGen’s code available on Github. This move aligns with the company’s commitment to transparency and sharing AI models, as seen with the release of open-source AI models like LLaMA earlier this year. In contrast, Google decided not to make MusicLM publicly available due to ethical concerns.
Felix Kreuk, one of the co-authors of the study, took to Twitter to explain how MusicGen works. The system is built on a single-stage transformerLM, eliminating the need for cascading multiple models. The researchers addressed potential limitations and ethical considerations, ensuring that all data used to train MusicGen was covered by legal agreements with music libraries like Shutterstock. They also acknowledged the dataset’s potential lack of diversity, with a larger proportion of western-style music.
Meta acknowledged the impact of generative AI on the music industry and expressed concern about its potential as unfair competition for artists. However, the company believes that open research and access to advanced controls, such as the melody conditioning introduced in MusicGen, can benefit both music enthusiasts and professionals. As music streaming services continue to grapple with AI and copyrighted music, Meta emphasizes the importance of equal access and responsible development.
Twitter users have started sharing music generated by MusicGen through the system’s demo website, showcasing the system’s potential and the interest it has generated among users.
Article by PRITESH PATEL