Google Now Produce AI That Serve Music

Putri Alivia Rifaih . January 30, 2023

Picture: Ai.Google

Techspace - It's certainly not unexpected that AI would go into music given how gradually it has entered every aspect of our digital life, from essay writing to discussions with therapists to the creation of unique art.

It's also not unexpected that Google, a significant player in the tech industry, would be the first to debut their AI.

Similar to how systems like DALL-E create visuals from textual prompts, Google engineers have created an AI that can generate minutes-long musical compositions from text prompts and even convert a whistled or hummed tune into other instruments (via TechCrunch). 

The business has published a few samples it created using the model, which is dubbed MusicLM, even if you can't play around with it for yourself.

MusicLM Capabilities 

Picture: Shutterstock

Creating music is a hierarchical sequence-to-sequence modeling challenge that results in 24 kHz music that lasts for many minutes.

In terms of audio quality and conformance to the written description, this approach outperforms earlier ones.

The capacity to produce melodically-driven music that is also texturally-conditioned.

The approach may be used with both text explanations and whistled or hummed tunes, which makes the final point particularly intriguing. It is now possible to combine the two input methods and deliver a multimodal prompt that includes both text and a hummed melody. 

As a result, MusicLM can then create music that is automatically changed into the text's style.

Additionally, Google made available the MusicCaps dataset, a collection of 5500 music-text combinations that will allow academics to better understand how MusicLM is generated.

With its capacity to make music that is both textually accurate and of a high calibre, MusicLM has the potential to completely change how we create songs and engage with music.

Even though MusicLM can replicate human vocals and seems to get the pitch and general sound of voices right, there is unmistakably an off-putting element to them. They sound grainy or staticky, which is the best way to put them. 

Examples of the results produced by the tool are displayed on the paper's website. Someone hums "Bella Ciao," a late 19th-century Italian folk song, in one case. The application then makes music with instruments in a variety of styles, including guitar solo, string quartet, and jazz with saxophone, based on the humming.

Google is being somewhat more cautious with MusicLM than some of its competitors may be with comparable technology, as it has been with prior excursions into this form of AI. The article ends with the statement, "We have no plans to disclose models at this point," noting dangers of "possible misuse of creative content" (read: plagiarism) and "potential cultural appropriation or distortion."

Songs are constructed using additional timed text prompt sequences, a sound library, and other AI prompts derived from sources like art archives. (Opens in a new window)

Examples of the AI-generated songs have already been uploaded to Google's Github account as part of a preliminary release of the dataset MusicCaps, which consists of 5,500 music-text pairs.

As of right now, Google AI music producer won't be available anytime soon due to persistent issues with cultural programming biases, bugs, and worries about plagiarism that need to be addressed before launch, according to the company.

Google Is Sued By The Justice Department For “Monopolizing” Advertising

teknologi id bookmark icon

Leave a comment