Friday , December 4 2020

Israeli researchers are creating artificial intelligence capable of writing personalized jazz solos

Author mr. students Nadav Bhonker and Shunit Haviv Hakimi, together with their advisor prof. With the early El-Yaniv at the Henry and Marilyn Taub School of Computer Science at the Technion-Institute of Technology, the paper suggests that it is possible to model and optimize personalized jazz preferences.
Learning to make music is an ongoing challenge in the world of artificial intelligence (AI). An even more difficult task is to create musical works that suit human specifics.

In the BebopNet project, Bhonker and Haviv Hakimi, both amateur jazz musicians, focused on a personalized, symphonic, monophonic generation of jazz improvisations limited by harmony.

To meet their goal, they introduced a three-step pipeline:

First, the researchers trained BebopNet, a musical language model, to be able to generate symbolic jazz saxophone improvisations for any chord progress.

To build their initial data set, the researchers used hundreds of original jazz solos performed by saxophone giants, including Charlie Parker, Stan Getz, Sonny Stitt and Dexter Gordon.

The article also presents a “plagiarism analysis” that compares all prominent musicians and BebopNet to assess the originality of the solo.

Second, AI begins by gathering a personal dataset for the user, training the personal preference metric to predict notes that reflect the user’s unique personal taste.

Each user is presented with jazz improvisations that they must evaluate according to their wishes, after which a regression model is used to predict the user’s taste.

Finally, the model uses a process called “air search” to optimize the note generation process to suit the specific taste of the user.

“While our computer-generated solos are locally coherent and often interesting or enjoyable, they lack the qualities of professional jazz solos related to general structure such as motif development and variation,” the authors said.

El-Yaniv said he hopes this challenge will prevail in future research. Preliminary models based on a smaller data set were significantly weaker and it is possible that a larger data set would make a significantly better model.

However, in order to obtain such a large set, it might be necessary to abandon the symbolic approach and rely on audio recordings, which can be collected in much larger quantities.

Source link