Google’s AI rewrites “Bella Ciao”. The music is about to change

Waiting for Sparrows – the artificial intelligence with which Google will respond to ChatGpt – the Mountain View giant has unveiled a new, interesting AI model capable of generate music starting from a simple text description.

And therefore, just as today we use ChatGpt to obtain a story or a poem starting from a written question, in the future we will be able to obtain the music we want by asking MusicLMfor example, to produce “the soothing melody of a violin accompanied by a distorted guitar riff”.

In the scientific paper released by Google, and signed by several Italian researchersthere are numerous examples of sounds and very short songs produced by MusicLM, which she has been trained with a volume of data that includes 280,000 hours of music.

There are, for example, audio produced starting from a very detailed description. Like this: “A fusion of reggaeton and electronic dance music, with a spatial and otherworldly sound. It invites you to lose yourself in space and evokes a sense of wonder and awe, while being danceable”. The thirty seconds generated by this text respond, almost to perfectionto the description provided through words.

In this regard, it should be remembered that generative AI – of any kind – to give its best needs to be guided as much as possible by the words of human beings. A superficial, generic or terse question, in fact, will always produce a rather trivial and poor quality content. The secret, therefore, lies in the so-called ‘prompts‘, i.e. the text command that is given to the artificial intelligence to obtain not only text, but also images (think of Give her And Midjourney) or, indeed, sounds.

With MusicLM, in addition to the text, you can use a pre-existing melody to affect each prompt. In the study published by Google there is an example of this technique applied to “Bella ciao”. The first ten seconds of this song are offered to the AI, simply whistled or hummed, and then specified with a text prompt the alternate version you intend to obtain. And so, magically, “Bella ciao” is performed by “an a capella choir”, or resounds in the form of a “guitar solo”, or again as the result of “a jazz group equipped with a saxophone”.

MusicLM’s creativity can produce unexpected results. Have you ever wondered what a painting sounds like? Well, Google researchers asked the AI, providing the text description of art masterpieces such as “The Persistence of Memory” by Salvador Dali. It was enough to enter, as a prompt, an excerpt from the entry dedicated to this painting from the Encyclopaedia Britannica, to obtain 30 seconds of an ethereal melody characterized by the notes of a piano.

But it doesn’t end there. The AI ​​can also be asked to set a musician’s experience level. “Play like a beginner pianist”, for example, or “like a professional pianist”. Or, in the extreme, as “a madly fast professional pianist”.

What MusicLM can’t do yet, on the other hand, is to produce a song with sensible verses, clear and distinct words. Everything comes together, when he tries, with an effect similar to the one obtained when an AI is asked to generate a caricature starting from one of our selfies: the result never completely resembles us, it is often the result of the union of faces taken from the database from which each model learns.

“MusicLM generates 24 kHz music that remains coherent for several minutes – reads the paper signed by Google researchers -. Our experiments show that MusicLM outperforms previous models in both audio quality and adherence to text description.”

Effectively, there are already AI models similar to MusicLM. One of these, Dance Diffusion, was developed by Google. But there are other extremely curious ones, such as Riffusion which allows you to generate music – incredibly – starting from an image, to be precise from a spectrogram. But none of these artificial intelligences, so far, had been able to reach the sonic complexity and high fidelity that distinguish MusicLM.

For now we’re only talking about a scientific paper, that is to say the study of a technology that is still in the testing phase and above all that it is not yet open to the publicbut the results of Google’s work are quite impressive and suggest, once again, how much AI will change (and simplify) the way we produce content designed for the masses.

Artificial intelligence

Can AI be used to make videos and make money on YouTube? We tried and we are not so convinced

by Pier Luigi Pisa

December 29, 2022

For text and images the revolution is already underway. We’re almost there on audio. And the videos too, soon, they can be generated quickly by an artificial intelligence.

But the starting point, the imagination behind each prompt, will continue to be ‘human’. In fact, only a man can wonder what soundtrack a painting can have.