Scholars at Google have created an AI system that can “create” high-fidelity music just based on text hints given by users.
Google’s AI tool used 280,000 hours of audio for training. The new system called MusicLM uses the AudioLM music generator. The latter is an AI-driven music creation tool that made some noise back in October.
Gizchina News of the week
It works in the following way: a user inputs the description of music in the form of text (1); then, the system produces the corresponding audio (2). For this, the system needs to know the type, length, genre, etc. Apart from these details, users can also provide more info with MusicLM. The latter can also create music that “induces the experience of being lost in space.” Another task of the tool could be the creation of music based on a melody that the user whistles or hums.
We should also point out that MusicLM has some bugs. For instance, it might create music that “remains consistent over several minutes.” But the AI system is still the best tool in the niche. So if users provide more details with it, the system will create a better melody.
The biggest highlight of this tool is a combination of multiple neural networks. Each of them can control a separate step in the music generation workflow. Instead, MusicLM uses just one. Google uses the so-called Transformer architecture. The company released it in 2017.
As you’ve realized, it considers a variety of data points when making a decision, such as how to produce a piece of music. The advantage of this system is the improved accuracy due to a bigger impact of data on the processing outcome.
Google hasn’t made the MusicLM code public yet. At the moment, researchers published an AI training dataset. It consists of about 5,500 musical compositions. Each of them has a text description to make it simpler for neural networks to understand.