Move over Google, Meta has finally joined the AI music generation race by releasing its own innovative tool called MusicGen. What sets MusicGen apart from Google’s offering is that it’s open source, allowing anyone to explore and experiment with its capabilities.
Last month, Google released its “experimental AI” tool, MusicLM, to the public after its initial unveiling in January. This tool allows users to generate music from text prompts or humming, presenting them with two versions of the requested song. Users can then vote on their preferred version, contributing to the ongoing improvement of the AI model.
With MusicGen, you can transform a simple text description like “An energetic ’80s pop song with booming drums and captivating synth pads” into a 12-second audio clip. If you want more control, you can even guide MusicGen with an audio reference track, giving it a melody to follow alongside the description.
Using prompts like “pop dance track with catchy melodies, tropical percussion, and upbeat rhythms, perfect for the beach” or “acoustic folk song to play during road trips, guitar flute choirs,” MusicGen can generate short music clips. Moreover, it can be steered by referencing specific eras or songs.
To train MusicGen, Meta leveraged an extensive dataset of 20,000 hours of music. This included 10,000 high-quality, licensed music tracks and 390,000 instrument-only tracks sourced from Shutterstock and Pond5, both major stock media libraries. While Meta hasn’t released the training code, they have made pre-trained models available for those with suitable hardware, primarily a GPU with around 16GB of memory.
Now, let’s talk about MusicGen’s performance. While it’s not yet ready to replace human musicians, it shows promising prospects.
For basic prompts like “ambient chiptunes music,” MusicGen produces reasonably melodic results that rival or even slightly surpass Google’s AI music generator, MusicLM. However, it’s important to note that MusicGen still has a long way to go before winning any awards.
But here’s where MusicGen shines: when faced with a more intricate prompt like “Lo-fi slow BPM electro chill with organic samples,” it surprises with its musical coherence. The output would easily find a home on Lofi Girl, known for its lo-fi music streams on YouTube.
As demonstrated by projects like Riffusion, Dance Diffusion, and OpenAI’s Jukebox, generative music is undoubtedly improving.
However, ethical and legal concerns still need to be addressed. AI systems like MusicGen learn from existing music to produce similar effects, which raises discomfort among some artists and generative AI users.
The rise of homemade tracks that utilize generative AI to replicate authentic sounds has led to viral sensations. Music labels have promptly flagged such tracks to streaming platforms, citing intellectual property concerns, and have generally prevailed in the ensuing legal battles.
Nonetheless, the issue of whether “deepfake” music infringes upon the copyrights of artists, labels, and rights holders remains unresolved.
Thankfully, guidance on these matters may be on the horizon. Several lawsuits currently in progress will likely shape the landscape of music-generating AI, including one that focuses on the rights of artists whose work is used to train AI systems without their knowledge or consent.
In Meta’s case, they have chosen not to impose restrictions on how MusicGen can be used.
They assure us that all the music used for training the model was obtained through legal agreements with the right holders, including a partnership with Shutterstock.
However, it’s worth noting that the music used for training MusicGen primarily came from media libraries like Shutterstock and Pond5, not industry giants like Universal Music Group or renowned artists like Taylor Swift.
Meta acknowledges the potential lack of diversity in their dataset, which contains a larger proportion of Western-style music.
As AI continues to advance rapidly, writers, artists, and musicians find themselves grappling with the transformative impact it will have on their creative processes.
Avenged Sevenfold frontman Matt Sanders (a.k.a. M. Shadows) is one artist who sees the potential of AI, emphasizing its usefulness for songwriters. He believes AI can expedite the generation of unique ideas, acting as a valuable tool to discover those hidden musical gems.
The use of generative AI in music has already sparked debates and made headlines. From Grimes advocating for “killing copyright” to famous artists like Oasis and The Weeknd encountering AI-generated replicas that raise questions about music licensing and content ownership, the impact is undeniable.
Even Paul McCartney is leveraging AI to create a never-before-released Beatles song.
As the world of AI music generation continues to evolve, we can expect further advancements, along with the necessary discussions and actions to address the ethical and legal implications. Meta’s MusicGen is just one example of how AI is shaping the future of music creation, providing both excitement and challenges for artists, listeners, and industry stakeholders alike.
If you’re curious about MusicGen, you can explore it on Github, where Meta has released it to the public. You can also try out its demo version via Hugging Face.