Introducing AudioCraft AI Music Generator
Meta has unveiled an open-source AudioCraft AI music generator that creates audio based entirely on a user’s text query. AudioCraft combines three separate AI models: MusicGen is designed to create music and is trained on “20,000 hours of music owned by Meta or licensed specifically for this purpose”, AudioGen generates sounds and environmental effects, and EnCodec provides quality audio processing.
Subheading 2: Shaping the Future of Music Generation
Musicians have been experimenting with electronic sound for a long time, but computer programs create music based on existing sound samples. The audio from AudioCraft is generated from the text prompt only. Meta provided journalists with audio samples generated with AudioCraft. Sound effects such as whistles, wind, sirens and car horns sounded very authentic. But the guitar parts seemed unnatural to the listeners.
Subheading 3: Challenges and Hopes for AudioCraft’s Potential
Right now, the music generated by AudioCraft is most reminiscent of muzak (a slightly derogatory term applied to most forms of background music, regardless of source, “elevator music”) or unfussy atmospheric ambient, and does not claim to be the next big pop hit. However, Meta believes that AudioCraft can usher in a new wave of musical fashion, just as early synths once did.
Meta has acknowledged the difficulty of creating AI models for music generation, a task that a company spokesperson claims is orders of magnitude more difficult than AI text generation like Llama 2. The company believes AudioCraft’s open source code will help diversify the data used to train it.
“We understand that the datasets used to train our models are not very diverse: most of the Western-style music, audio-text pairs with text and metadata are written in English,” explained a Meta representative. “By sharing the code for AudioCraft, we hope it will be easier for other researchers to test new approaches to limit or eliminate the potential bias and misuse of generative models.”
Meta is far from being a pioneer in AI-assisted audio generation. Google’s large MusicLM language model generates audio quite successfully, although it is only available to researchers. An AI-generated song with voice resemblance to Drake and The Weeknd went viral instantly. Recently, Grimes allowed the use of imitation of her voice in AI tracks, reminds NIX Solutions. In turn, record labels and artists have already sounded the alarm, as many AI models can use copyrighted materials for training.