Meta has introduced AudioCraft, an open-source AI code that allows users to create music and sounds through generative AI.
Three AI Models for Sound Generation
AudioCraft consists of three AI models: MusicGen, AudioGen, and an improved EnCodec decoder. Each model serves a different purpose in sound generation.
MusicGen: Text-Based Music Generation
MusicGen uses text inputs to generate music. The model was trained on 20,000 hours of music specifically licensed for this purpose.
AudioGen: Audio from Written Prompts
AudioGen creates audio from written prompts, simulating various sounds like barking dogs or footsteps. It was trained on publicly available sound effects.
Realistic Audio Samples
AudioCraft's sample audio, including whistling, sirens, and humming, sounded remarkably natural. However, certain elements like guitar strings still felt artificial.
The AI Music Landscape
Meta isn't the only player in the AI music game. Google's MusicLM and other AI-generated songs have also garnered attention. Musicians like Grimes encourage AI-made songs.
Elevator Music vs. Pop Hits
Currently, AudioCraft might find more use in creating elevator music or stock songs for ambiance rather than producing chart-topping pop hits.
A New Wave of Music
Meta envisions MusicGen as a new type of instrument that could revolutionize music, similar to how synthesizers changed the game once they became popular.