Google has developed an AI capable of generating music

Google has Developed an AI capable of Generating Music

Google has developed an AI capable of generating music

Google has Developed an AI capable of Generating Music

Google has developed an AI capable of generating music

Google unveiled, in an article published on January 26, 2023, Google has developed an AI capable of generating music the underside of “MusicLM”, a new artificial intelligence allowing to compose any type of music from simple textual indications.

After ChatGPT, anw artificial intelligence has just emerged. It allows this time to compose whole pieces of music from a simple phrase. Back to this announcement signed by Google.

With its hugely popular text and image generators ChatGPT and DALL-E, already widely used, the OpenAI company has taken a step ahead in the race for the most powerful artificial intelligence. But Google has not said its last word on the matter. In a research article published on January 26, the company unveiled MusicLM, an algorithmic model for producing songs from simple textual indications, such as “soothing violin melody” or “main soundtrack of a arcade game”.

Although the subsidiary of the Alphabet group does not intend to distribute its bot in the very short term, the company has already put online a series of musical samples generated using MusicLM (melodies on the piano, relaxing jazz pieces, etc.). Whether the text descriptions are simple (“rock song”) or more detailed (“1990s Berlin techno with a bass and a powerful kick”), the AI developed by Google reaches a level of perfection hitherto never reached.

Among the examples shared on its Github page, Google offers a selection of pieces composed from table captions, including Munch’s cry and Matisse’s dance. Another string to its bow, the AI can generate music inspired by places (“Caribbean beaches”) and times (“the 1950s”), but also transform melodies whistled or hummed according to the style described. in the caption. Generating 24kHz music that stays consistent over several minutes, “MusicLM outperforms previous systems in terms of audio quality and adherence to text descriptions,” Google researchers boast.

With MusicLM, Google confirms its ambitions in terms of artificial intelligence research. In line with the “red code” issued in December 2022 in order to compete with the competition in the field, the company accelerated the launch of around twenty new products, as well as a new version of Google Search equipped with functionalities. artificial intelligence, according to information from the New York Times. Among these innovations are the text generator Sparrow, the chatbot-doctor who answers medical questions MedPaLM, and therefore the newly unveiled MusicLM project.

Read also: Artificial Intelligence: definition, advice, comparisons and Examples of AI technology

Industry Competition: How to Analyze and Stay Ahead of Your Competitors

From Illiac Suite to iRap

In order to produce the approximately 5,500 “music + text” combinations available today, Google trained MusicLM from 280,000 hours of royalty-free music available on the Free Music Archive platform. To work, the AI takes bits of existing sound and matches them to words. The system builds on an earlier AI model called AudioLM, introduced by Google in October 2022, as well as other services such as SoundStream and MuLan.

Last December 2022, an amateur AI project called Riffusion was already able to generate music from textual descriptions, without guaranteeing fidelity to legends. Other attempts have been led by Harmonai and its bot Dance Diffusion, which boasts of being “the first in a suite of generative audio tools for producers and musicians”, or OpenAI’s Jukebox, “a network neural that generates music, including rudimentary vocals, as raw audio in a variety of genres and artist styles”. But due to technical limitations and limited training data, none were able to produce songs that were particularly complex compositionally or truly faithful to the textual cues submitted.

While advances in artificial intelligence in music are focused on the last ten years, generating tracks through technological devices is nothing new. The Illiac Suite (later retitled String Quartet No. 4), composed in 1957 by the American Lejaren Hiller for his String Quartet No. 4, is widely considered to be the first score generated by an electronic computer. Forty years later, it was David Bowie who made the AI application Verbalizer famous, by randomly generating new lyrics for his songs. Even more recently, Sony has developed Flow Machine, an artificial intelligence capable of generating songs in the style of great composers. In 2016, the Japanese company notably released Daddy’s Car, a single inspired by the style of the Beatles. Since then, music lovers have been constantly inventing new algorithmic models, such as iRap, an AI capable of generating rap or slam lyrics.

Start up Character.AI begins to overshadow ChatGPT as a Game-Changing AI Conversational Platform

Avoid plagiarism

Whether it’s for the Flow Machine AI or more recently for Google’s MusicLM project, the age-old debate around plagiarism comes to the surface every time a new music AI is unveiled. Because MusicLM, as for the image generators DALL-E, Midjourney or StabilityAI, bases its automatic learning on already existing music. For the time being, none of the pieces used as a database by Google for its AI are subject to copyright, unlike its competitors specializing in the generation of images, some of which are also the subject of a complaint. for plagiarism.

Aware of this problem, the digital giant has decided not to make MusicLM accessible to the general public. “We recognize the risk of potential diversion of creative content associated with the use case, write the co-authors of the article published on January 26. We strongly emphasize the need for more future work to address these risks associated with music generation.” Currently, researchers have found traces of existing, identifiable songs in approximately 1% of MusicLM-generated tracks, which opens the door to potential copyright infringement.

In addition, Google wishes to guard against programming biases that may lead to a lack of cultural representativeness: “The samples generated will reflect the biases present in the training data, raising the question of the adequacy of the music generation for the underrepresented cultures in the training data, while also raising concerns about cultural appropriation,” the researchers note.
It will therefore be necessary to wait several months, even several years, before MusicLM is accessible. Until then, Google intends to improve its technology, in particular by facilitating the recognition of negative turns in the textual indications allowing to generate music.

AI and music: legal precedents for plagiarism?

In 2020, Jay-Z’s record company filed a copyright infringement lawsuit against YouTube channel Vocal Synthesis for using AI on the singer’s covers of songs like “We Didn’t Start the Fire” by Billy Joel. After initially removing the videos in question, YouTube reinstated them, saying the takedown requests were “incomplete”.

Sources: Yahoo, MusicTech

Photo credit: SpiritBunny via Pixabay

ChatGPT Flaws | Threatens to put several professions out of work, this time in the writing sector

Leave a Reply

Your email address will not be published. Required fields are marked *

The reCAPTCHA verification period has expired. Please reload the page.

AB Consulting: Elevate Your Business with Tailored Growth Strategies

Maximize your business potential with AB Consulting’s customized solutions in organizational development, stakeholder management, and leadership development. Achieve sustainable growth and excellence with strategies designed for your unique challenges.

Social Links