Meta* introduced AudioCraft, a music and sound generator based on text description
Miscellaneous / / August 03, 2023
A neurosynthesizer for any task: from simple noise to complex melodies.
Meta* has released a new open source tool called AudioCraft. It allows users to create music and sounds entirely with generative AI.
AudioCraft consists of three models, each dealing with different areas of sound generation:
- MusicGen allows you to create melodies by text prompts. She was trained on "20,000 hours of music owned by Meta* or licensed specifically for this purpose".
- AudioGen creates certain sounds from text prompts, imitating, for example, barking dogs or human steps.
- encodec allows you to process sounds, reducing the number of artifacts.
The company allowed members of the media to listen to some audio samples created with AudioCraft. The generated whistle and noise of the sirens sounded quite natural, but the sound of the guitar strings still seemed artificial, noted at The Verge.
The development of Meta* is not the only one of its kind. Google has already introduced MusicLM, a large language model that also generates sounds based on text prompts.
Record labels and artists did not appreciate such innovations. Many of them have already spoken out about the dangers of AI, as some models may use copyrighted materials for training.
AudioCraft on GitHub →
*Activities of Meta Platforms Inc. and its social networks Facebook and Instagram are prohibited in the territory of the Russian Federation.
Read also🧐
- 105 services based on neural networks for a variety of tasks