If image and video generation services have greatly increased in recent months, it is clear that few services dedicated to audio – and more particularly to Music – have seen the light of day until now. However, it’s not for lack of trying. Starting with the technological giants who nevertheless have all the keys to achieving this. Thus, Google – via its DeepMind branch dedicated to AI – and YouTube have worked together on a service for creating music generated by artificial intelligence.
Called Orca, the tool would be able to take a few simple instructions like lyrics, artist and genre, and generate something that sounded incredibly authentic, according to employees who tried it. The model behind this tool was notably trained on copyrighted music from YouTube.
Google “stupidly” followed in the footsteps of OpenAI
However, the platform not only prohibits the use of its videos for applications “independent”but also to access its videos by “automated means (such as robots, botnets or scrapers)”. Google, however, authorized this for a simple reason: after learning that OpenAI, using its speech recognition tool Whisper, literally sucked the audio from millions of hours of videos on YouTube in order to transcribe them and make them usable data for training its large language models, the Mountain View giant has decided to follow suit.
The firm has also protected itself from any accusation since its rules authorize it to exploit YouTube user data in order to develop new features for the video platform. However, it is unclear whether Google can use YouTube data to create a commercial service beyond the video platform, such as a chatbot.
A project abandoned due to the associated legal risks
Business Insider reveals that Google subsequently approached some record labels to release the Orca tool to the public, offering a revenue-sharing deal for the music and artists that Orca had trained on. The record companies hesitated, forcing Google to slow down the project, according to a source interviewed by our colleagues, adding that it was a “huge legal risk”. It is a safe bet that the firm would, despite its considerable power, have been a prime target of regulators if it had effectively trampled on its own rules to remain competitive.
If this project ultimately fell through, Google nevertheless launched other products, certainly less impressive, but similar. In May 2023, the firm unveiled MusicLM. It is a generative AI tool capable of generating music following a textual description. This may include instructions on the instruments used, the desired tempo and even the emotion to be expressed. The latter was trained on “a large dataset of unlabeled music”, according to the scientific publication published by the Google team.
Most recently, in November 2023, DeepMind announced a music generation AI model named Lyria, which was a scaled-down version of Project Orca. Users can ask the tool to generate music using the voice and musical style of certain artists who had explicitly worked with Google on the project, such as John Legend, although this is much more limited in scope than Orca .
Ex-employees continued the project via their own start-up
Some employees who worked on Lyria and Orca left the company to found a startup called Udio, which creates an AI music creation app. Based in New York, the start-up is the result of the work of several former Google DeepMind researchers.
It is backed by leading venture capitalists such as Andreessen Horowitz, but also artists such as will.i.am and Common, entrepreneurs – including Kevin Wall – and Instagram co-founder Mike Krieger. Its app, released to the public in April, was wildly successful in part thanks to the tool it used to create “BBL Drizzy,” a parody track tied to the feud between artists Kendrick Lamar and Drake.
The young company, however, was caught up by the record companies: last June, it found itself – just like the start-up Suno – pursued by the major record companies Sony Music, Universal Music Group and Warner Records. The latter, through the Recording Industry Association of America, filed two complaints against these start-ups for copyright infringement. The affair could be expensive: up to $150,000 per song copied.
Selected for you
