OpenAI has recently seen a lot of success with ChatGPT's stunning capabilities, including writing academic abstracts, planning a holiday party, and even writing a song in the style of Australian singer-songwriter Nick Cave. So much so that ChatGPT has Google worried about the future of its search engine, prompting the company to develop an AI version of Google Search alongside other AI projects, such as image-creation technology, per The New York Times. It's no surprise, then, that Google is venturing into yet another uncharted territory: AI-generated music. The search giant has built an AI bot that can create music from text descriptions or even stories reflecting a certain epoch.

Google refers to this AI system as "MusicLM," presumably its take on ChatGPT in terms of music creation. In a research paper published on January 26, Google researchers described MusicLM as a "model generating high-fidelity music from text descriptions." The system can produce music at 24 kHz, and it stays at that frequency for the next few minutes (via TechCrunch).

For MusicLM to create music, a highly detailed prompt noting the styles or genres must be written first. Captions like "the main soundtrack of an arcade game...with a catchy electric guitar riff" or "a fusion of reggaeton and electronic dance music, with a spacey, otherworldly sound" can produce awe-inspiring results.

The AI can also produce longer soundtracks with short text prompts, such as "melodic techno" or "relaxing jazz," which last up to five minutes. It can even transform a short narrative or painting description into music. For example, it has created a fascinating melody based on Jacques-Louis David's "Napoleon Crossing the Alps" and Edvard Munch's "The Scream."

Google has trained MusicLM on existing copyrighted musical content, so it's no wonder that around 1% of examples can be associated with existing songs, the researchers noted. This could raise concerns about MusicLM's ability to easily lift from existing copyrighted materials to create its own song.

As a result, Google isn't planning on releasing this model to the public anytime soon. Having said that, the company has made 5,500 music-text pairs available to anyone who wants to see — and hear — how MusicLM does its trick. This dataset, known as "MusicCaps," is now available on Google's Github page and shows the text prompts and resulting audio side-by-side.