AudioCraft
lastmod 2023-08-03created 2023-08-032 edits
Meta releases generative AI for audio as open source
Features
- Generative AI that generates high-quality audio and music from text-based user input
- Composed of MusicGen + AudioGen + the EnCodec decoder, responsible for music generation, audio generation, and audio compression respectively
- Can be used for research purposes and to help people better understand the technology
- Lets professional musicians compose new pieces without playing a single note on an instrument, or indie game developers implement realistic sound effects and background noise at low cost
Hacker News comments
- Open-sourcing AudioCraft: Generative AI for audio | Hacker News
- Generative AI models like AudioCraft used in advertising make A/B testing easy and remove the need for creative staff.
- Meta distinguishes itself from OpenAI by emphasizing that it uses patented data and does not use data it has not paid for.
- The CC-BY-NC license used by AudioCraft is not considered an open source license under the Open Source Definition.
- There are concerns that the past may disappear under the flood of AI-generated content, and that it may become hard to tell real information from fake.
- There is speculation about the ability to generate music based on existing music libraries.
- Some people believe that copyright holders controlling the data input to AI models could lead to a "thought-policed" society.
- It is unclear why Meta chose to open-source these models.
- Some users are having trouble finding documentation on installation and how to use the models.
- There is a noticeable difference in audio quality between AudioCraft's MBD and EnCodec variants.
- One commenter suggests that Spotify could use generative AI models like AudioCraft to generate its own music and reduce its dependence on labels for profitability.