In a bid to meet up with its Massive Tech friends within the AI arms race, Meta has been unveiling a slew of AI instruments, the most recent of which is Audiocraft, an AI software that may generate audio and music from textual content prompts.
AudioCraft consists of three fashions: MusicGen, AudioGen and EnCodec. MusicGen was skilled on roughly 400,000 recordings together with textual content description and metadata, amounting to twenty,000 hours of music owned by Meta or licensed particularly for this function. It generates music from textual content prompts, whereas AudioGen, which was skilled on public sound results, generates audio from textual content prompts.
Immediately, Meta launched an improved model of the EnCodec decoder, which permits higher-quality music era. Concurrently, the corporate is launching its pre-trained AudioGen fashions, enabling customers to create an array of ambient sounds and auditory results reminiscent of a canine’s bark, automotive horns, or footsteps on wood surfaces. Moreover, Meta is making the entire set of AudioCraft mannequin weights and code accessible to the general public.
These fashions might be open-sourced, permitting researchers and practitioners to coach their very own fashions with their very own datasets. Based on Meta, The AudioCraft household of fashions is able to delivering high-quality audio, whereas remaining user-friendly.
“We see the AudioCraft household of fashions as instruments for musicians’ and sound designers’ skilled toolboxes in that they’ll present inspiration, assist folks shortly brainstorm, and iterate on their compositions in new methods,”
Meta wrote in a weblog submit.
AudioCraft serves as a unified platform encompassing music, sound, compression, and era, all inside a single framework. People aiming to construct higher sound mills, compression algorithms, or music mills can accomplish that inside the similar code base, constructing upon the inspiration laid by others within the subject.