In a bold move to keep up with its tech contemporaries, Meta brings a riveting transformation to the world of music through the power of artificial intelligence. Breaking barriers, the tech giant has open-sourced its AI-powered music magician, named MusicGen.
This intriguing tool is capable of spinning melodious magic. Just give it a simple textual description - maybe something like "An '80s driving pop song with a strong drums base and synthesizer pads creating a backdrop" - and voilà! MusicGen can conjure about 12 seconds of tuneful delight from this prompt. You can even guide this musical wizard with a reference audio, maybe a favourite song, and it will attempt to mimic the description and melody.
We present MusicGen: A simple and controllable music generation model. MusicGen can be prompted by both text and melody.
— Felix Kreuk (@FelixKreuk) June 9, 2023
We release code (MIT) and models (CC-BY NC) for open research, reproducibility, and for the music community: https://t.co/OkYjL4xDN7 pic.twitter.com/h1l4LGzYgf
MusicGen owes its prowess to a vast training database encompassing 20,000 hours of music. This includes 10,000 exceptional licensed tracks and nearly 400,000 instrumental-only melodies from high-profile stock media libraries like Shutterstock and Pond5. However, Meta hasn't shared the training code but graciously provides pre-trained models for anyone with suitable hardware, mainly a GPU boasting around 16GB memory.
So, does MusicGen hit the right notes? Generally, it strikes a harmonious balance. It doesn't spell doom for human composers but demonstrates a decent competency. The generated tunes from prompts like "ambient chiptunes music" are appealing and arguably a notch above Google's MusicLM.
While the sphere of generative music is making strides with creations like Riffusion, Dance Diffusion, and OpenAI’s Jukebox, it also treads the fine line of ethical and legal issues. The AI learning from existing music to weave similar effects isn't sitting well with everyone, particularly some artists and AI users.
Rapidly, tracks crafted by generative AI mimicking authentic sounds are gaining traction. Quick on their feet, music labels are raising flags for potential intellectual property breaches, generally succeeding in their claims. Still, the domain lacks definite clarity on whether this "deepfake" music breaches copyright rules.
The murky waters might soon clear as lawsuits inch closer to verdicts, potentially influencing the future of music-generating AI. One critical case focuses on artists' rights when their works unknowingly train AI systems.
Regardless, Meta ensures that MusicGen abides by legal agreements with all right holders, including Shutterstock, without imposing any usage restrictions. All eyes are now on this open-source melody maestro, promising to redefine our musical experiences.
