The appearance of ChatGPT It has been one of the sensations of these last months. This new artificial intelligence, developed by OpenAI, has conquered millions of users thanks to its amazing power to hold conversations, create codes or solve moral dilemmas. And Google already has the AI ready that can take some of the prominence away from ChatGPT.
The technological giant, as it has collected Tech CrunchHas published a new article about a text model que can generate high-quality music of any genre from a short textual description. Google has dubbed its new artificial intelligence MusicLM.
Hifi music from text descriptions
This music creation tool, as explained by Google, is made up of SoundStream, W2v-BERT and MuLan, which are organized in what they call a sequence-by-sequence hierarchical training task. And the results that have been achieved with MusicLM are really good.
The company has ensured that a small explanation is enough to obtain a good result. It is enough to ask for “a relaxing violin melody accompanied by a distorted guitar riff” and MusicLM starts generating music “at 24 kHz that stays constant for several minutes”.
He may also be asked to transform the pain of Pablo Picasso’s Guernica into a melody, to create a piano piece played by a novice, or something that goes well with a 1950s disco. And he will do it with great precision. But that is not all. MusicLM can even “transform whistled and hummed melodies according to the style described in a footer.”
Other similar tools such as Riffusion, AudioML, Jukebox or Dance Diffusion had already begun to explore this territory, but they were halfway due to having several limitations when it comes to creating musical pieces of a certain complexity. In this sense, the developers have guaranteed that the new AI “exceeds previous systems both in audio quality and in adherence to textual description.” Although, as they have recognized from Google, it is far from being a perfect artificial intelligence.
It raises many ethical issues.
To arrive at these results, the artificial intelligence has been subjected to training with an archive of 2,80,000 hours of music to learn how to generate coherent songs for descriptions of “significant complexity”.
Therefore, the tool that Google has recently presented could revolutionize the music industry. I could, conditionally. Because it won’t. At least for now and if Google does not change its mind. “We have no plans to publish models at this time,” says the document published by Google. The company has alleged that there is a risk of plagiarism of creative content, cultural appropriation or misrepresentation.
During a trial, Google researchers found that approximately 1% of the music MusicLM generated was an exact imitation of the songs it had trained on. It may seem that this percentage is not too much, but it is a high enough level that they prefer to save the AI until it is fully prepared.
Google’s caution with MusicLM contrasts with how lightly other companies have released their AIs, even though they are buggy, unfinished, or otherwise problematic. Meta has already complained about this recently, specifically about ChatGPT and through Yann LeCun, vice president and chief scientist for Artificial Intelligence at the company led by Mark Zuckerberg. According to LeCun, ChatGPT is not a breakthrough because it has too many bugs, in addition to vampirizing the information found on the Internet.
Despite not even having planned the launch of MusicLM, Google has made more than 5,000 combinations of music and text available to users so that they can appreciate the power of artificial intelligence. And only those developers of other similar musical models will be able to use these pairs to be able to train and evaluate their respective tools, although they will not be able to publish the musical pieces obtained.
We want to thank the author of this write-up for this awesome web content
Google creates an AI that can revolutionize the music industry, but is too scared to publish it