New Stable Audio neural network generates music based on text description
Miscellaneous / / September 14, 2023
From the creators of Stable Diffusion.
Stability AI, known mainly for neural networks for generating pictures (Stable Diffusion, Stable Doodle and not only), released new neural network Stable Audio. As the name suggests, it generates audio clips.
Everything works in much the same way as image generators based on text descriptions. The user specifies keywords (for example, "melodic lo-fi hip-hop, melodic, 85 BPM" or “death metal with powerful guitar riffs and fast drums”), the desired duration and waits for the result.
We used the same model as in Stable Diffusion as a basis, but trained it on musical compositions instead of images. In total, she was fed about 800 thousand tracks from the stock music site AudioSparx - or about 19,500 hours of different sounds.
The creators note that the main feature of Stable Audio is the ability to generate compositions of a given length. Previously, neural networks only worked with a fixed duration: if they were trained on 30-second audio clips, they could only generate 30-second compositions. To be able to adjust the duration of a track, the developers had to change the model and add metadata for the beginning and end of the composition.
Stable Audio is offered in three models. The free version allows you to generate no more than 20 songs lasting up to 45 seconds per month. There's also a Professional subscription that lets you create up to 500 tracks up to 90 seconds long for $12 per month (≈1,200 rubles) and the Enterprise option for companies with the ability to select the generation volume and price individually ok. You cannot use the generated music for commercial purposes without a paid subscription.
As with other similar neural networks, Stable Audio is aimed more at content creators than professional musicians. Such tools are suitable for quickly creating background music for podcasts and videos when you don't have the time or budget to collaborate with a composer. It can also replace stock sounds if you want unique laughter or crowd noises.
You can try Stable Audio on the official website. You will need to register or log in with a Google account. At the start, there may be interruptions in access due to the heavy load on the server.
Stable Audio →
More new neural networks🦾✨
- The AIDA virtual assistant from Sber will diagnose patients in Moscow clinics
- New app Artisse generates cool photos with the user's face
- Stability AI introduced the chatbot Stable Chat - a free analogue of ChatGPT