However, the company acknowledges the challenges in creating AI models for music, as audio contains millions of action points compared to written text models. To address the lack of diversity in the training datasets, Meta has open-sourced AudioCraft, hoping other researchers can help limit or eliminate potential bias and misuse. The move comes amid concerns from record labels and artists about AI models using copyrighted material for training.
Key takeaways:
- Meta has released an open-source AI model called AudioCraft, which allows users to create music and sounds entirely through generative AI. It includes three models: MusicGen, AudioGen, and an improved version of Meta’s EnCodec decoder.
- Despite the progress in AI-generated music, the sounds still feel somewhat artificial. Currently, AudioCraft's output is more suitable for background music rather than chart-topping hits.
- Meta acknowledges the challenges in creating AI models for music, as audio contains millions of data points compared to text models. The company has open-sourced AudioCraft to diversify the data used for training and to limit potential bias and misuse.
- There are concerns in the music industry about the use of copyrighted material for training AI models, and the potential legal implications. The industry has a history of litigation over copyright infringement.