Google develops new AI system for generating high-fidelity music

Google LLC researchers have developed a synthetic intelligence system that may generate high-fidelity music primarily based on a textual content description supplied by the consumer.

Google detailed the system in a Jan. 26 research paper noticed in the present day by TechCrunch. The AI, often called MusicLM, was skilled on 280,000 hours of audio. It’s primarily based on an earlier AI-powered music generator referred to as AudioLM that was detailed final October. 

The brand new MusicLM system takes a pure language description of a musical observe as enter and robotically generates corresponding audio. Customers can specify the sort and variety of devices that the AI ought to simulate, the style and different particulars. 

MusicLM additionally permits customers to explain a observe in additional summary phrases. Throughout one inner check, Google researchers instructed the AI to generate music that “induces the expertise of being misplaced in area.” Furthermore, MusicLM is able to producing music primarily based on a melody whistled or hummed by the consumer.

The system generates music that “stays constant over a number of minutes” in some circumstances, Google’s researchers detailed. Inner checks decided that the AI system delivers larger audio high quality than current AI-based music turbines. Furthermore, it does so whereas adhering extra intently to the outline supplied by the consumer.

MusicLM includes not one however a number of neural networks that every handle a unique a part of the music technology workflow. The system’s neural networks are primarily based on the so-called Transformer structure. Launched by Google in 2017, the structure is a well-liked technique of designing AI methods that’s notably extensively used for pure language processing.

Neural networks normally analyze a number of knowledge factors when making a choice, akin to how a bit of music must be generated. The Transformer structure permits a neural community to prioritize the info factors it analyzes primarily based on their significance. An important particulars affect the processing outcome to a larger extent than the remaining, which improves accuracy.

The MusicLM system additionally incorporates an AI method often called sequence-to-sequence modeling. The method entails turning a bit of textual content, such a consumer’s description of a musical observe, into an summary mathematical illustration referred to as an embedding. This embedding might be changed into one other sort of information, akin to audio, extra simply than the unique textual content description.

Google has not but launched the code for MusicLM. Nevertheless, the corporate’s researchers revealed an AI coaching dataset to assist additional analysis into automated music technology. The dataset includes about 5,500 items of music that every embody an accompanying textual content description designed to make them simpler to interpret for neural networks.

Picture: Google