Artificial intelligence continues to enter the domain of art. And with a versatility worthy of the best masters. We have seen the bots philosophize, generate designs with more help than a few indications or write chronicles, sometimes, yes, with unequal luck. Y eso entre un etcétera cada vez más largo y complejo. In this context, Google has just presented its new development, MusicLM, a model capable of generating music from written instructions.
His approach is not entirely new. Other AI systems have already been presented centered in the music, such as Riffusion, Dance Diffusion, Jukebox or including AudioML, from Google itself. By the way, the music generated by IA has its own Eurovision, in which Spain has not been left behind.
What makes MusicLM relevant are its results -shared by Google and which you can consult in detail, with audio clips, on its web-, based in turn on a training feed with search 280,000 hours of melodies.
“Superior to the previous systems”
“MusicLM is a model that generates high-fidelity music from text descriptions such as ‘a relaxing violin melody backed by a distorted guitar riff'”, explain the authors of the article, in which they claim that their model “superior to the previous systems” both in terms of the quality of audio and in its capacity to adjust to the indications.
“We also demonstrated that MusicLM can be conditioned as much to a text as to a melody, and that it can be transformed whistling and humming pieces según el estilo descrito en un pie de texto”, zanjan. The article is available in Arvix.
What do their creators show? Its enormous versatility to generate melodies based on written instructions. In his list of examples are included pieces composed from guidelines such as “the main soundtrack of an arcade game” or “a fusion of reggaeton and electronic dance, with a spatial and other world sound”. Game orientations that can then be completed with others about rhythms, instrumentation, repetitions, details of persecution or development.
Do you want to know how it sounds? Here you have the result.
AI Text-to-Music has arrived.
MusicLM is a model by Google Research that generates high-fidelity music from text descriptions. Basically just enter some text and it will create the music. 🤯
Attached video shows sample text prompts and AI generated music. pic.twitter.com/C2mII0M5zv
— Dave Lee (@heydave7) January 27, 2023
Google has also shared other results generated from different, more abstract, detailed or even generic descriptions. In his list of audios, for example, melodies created with a sequence of orders are included, which leads to tunes that they tell a story, just like the soundtrack of a movie. An example? A well-cohesive piece, with sharp cuts and one minute long, that MusicLM developed from this succession: “Hora de meditar, hora de despertar, hora de corre y hora de dar el 100%”.
Another of the tests consisted of describing paintings by Salvador Dalí, Jacques-Louis David or Matisse, among other artists, in order to generate melodies with what those texts collected. The passages were taken from encyclopedias, specialized websites or even Wikipedia. Does this mean that everything in MuscicLM is perfect? No. There are compositions that sound distorted and when they resort to human voices – something for what is prepared, a priori – they are usually incomprehensible.
From now on, you will have to comply with the tests carried out by Google’s own experts. Techcrunch points out that Google is not planning to launch the model. At least in immediate form. Beyond its technical networks, the truth is that MusicLM poses challenges of an ethical character equal to those of the devil.
Perhaps the most thorny question of all concerns the rights of the author of the samples with which the model is trained and of those which are then surte to generate songs. During their investigation, the experts proved that more or less from 1% of the compositions generated for MusicLM were reproduced directly from already existing pieces with which he had trained.
“We discovered that only a small fraction of the examples were memorized accurately, while in 1% we identified an approximate coincidence. We insist on the need to continue working in the future to face the risks associated with the generation of music.”, acknowledge. We are left, at least for a moment, with the sample of melodies that MusicLM has already generated.
Cover image: Possessed Photography (Unsplash)
