Exploring MusicGen by Meta: A Review

I recently explored the capabilities of Meta’s MusicGen, an AI music generation tool, using the code available on GitHub.
The videos above showcase the results generated using the Large, Medium, and Small models.

Prompt and Setup

For my experiment, I used the following prompt: “Create a lo-fi, calm piece of music suitable for studying. It should be slow and quiet, ideally incorporating the sound of rain.”
The goal was to generate relaxing study music with a soothing rain backdrop.

Results and Observations

While the tool produced interesting results, there were several limitations:

• Duration Constraints: MusicGen only allows music length adjustments in 10-second increments.
• Time Limitation: The maximum music length supported is 2 minutes. Additionally, the music abruptly cuts off after 2 minutes, rather than concluding smoothly.
• Sound Quality: The audio quality felt somewhat low, which detracted from the overall listening experience.

Performance and Feasibility

I considered the possibility of improving these aspects through custom coding. However, there are practical challenges:

• Long Processing Times: Generating a 2-minute track on a T4 GPU in Colab took approximately 10 minutes. This extended processing time is not feasible for extensive testing or regular use.
• Cost Concerns: Although open source tools can save costs, the lengthy generation time ultimately makes it costlier due to higher computational demands.

Conclusion

Given these issues, it seems best to wait for future updates to MusicGen. The current version shows promise but requires significant improvements to be practical for regular use.
Stay tuned for more updates and reviews of AI-generated music tools!