|
Written by: Holly Zheng '22 Edited by: Shailen Sampath '20 In the final season of the 2018 comedy TV series “Mozart in the Jungle,” the protagonist, a conductor of a prestigious orchestra, hears about a mysterious composer who has just completed Mozart’s “Requiem,” Mozart's final work that he did not finish before his death. The news claims that this mysterious composer finished the piece exactly how Mozart would have done it. When the maestro later meets the composer, he is shocked and confused -- the composer turns out to be an AI-powered robot. Although this scene is from a fictional TV show, the reality of some recent technological advancements is not too different from it. On March 21st this year, Google celebrated the birthday of Johann Sebastian Bach by creating an interactive Google Doodle that took user-generated melodies and automatically harmonized them in Bach’s style. [1] The algorithm behind the doodle had studied 306 Bach chorales to apply to the characteristics of his music to new melodies. Many users even composed their own Bach-style chorales based on melodies of different genres, such as jazz and pop music. [2] The Google Doodle made it sound like Bach was alive once again. Music is one of the more creative and sentimental human activities that artificial intelligence has been trying to model. For example, one recent research proposed a new technique for music genre transfer. This model, based on the image rendering architecture, has allowed the shifting of a painting into a specific artist's style or changing a photo’s background from summer to winter. [3] Another research topic examined classification methods for separating singing voice from its instrumental accompaniment in a recording. This method adapted existing models on pixel classification from the computer vision field. [4] Some other breakthroughs involve more precise pitch estimation from recordings. [5]  These algorithmic models have made music production and education much easier. The advancements in music modeling provide tools for musicians to edit their melodies more precisely, perhaps even more so than experienced music virtuosos. Educational softwares also utilize artificial intelligence architecture to tailor exercises for each student. For a student displaying difficulty in rhythmic patterns, for example, the software might then generate “easier” rhythmic exercises to let the student build their skills gradually -- the definition of easiness would then depend on the student’s past exercise record extracted and analyzed by the model inside the software. Graph from Pitch Estimation paper [5]. The gradient represents pitch estimation accuracy, and the horizontal axis shows different musical instruments tested and their average frequency. Sometimes, it even seems that AI can replace musicians all together in a performance. There exist softwares that can take in a few sheets of music, or even an entire conductor score, and play back the melodies accordingly, no matter how complex the harmonic progression or the instrumentation is. In summer of 2017, a robot conductor named “YuMi” made its opera debut in Italy. [6] Although the robot consisted of only two arms that mimicked a human conductor instead of a whole body, it precisely conduct the rhythmic patterns of the piece. Musicians that performed with the robot conductor on stage in Italy, however, commented that there was little room for improvisation, because everyone had to follow the robot’s metronomic conducting. Under a human conductor, on the other hand, musicians usually have the opportunity to shape their music to their own interpretation. One of the musicians under the baton of YuMi said that it was a cool effect to see a robot conductor, but he didn’t believe that this would be the future, because there was little room for creativity. [6]  YuMi sharing an opera stage with Andrea Bocelli in Italy in 2017 Perhaps this reaction toward a robot conductor could predict the future of the AI music industry. The beauty of algorithm lies in its tireless training process and relentless precision. Machines outpower humans in these capacities, but so far they shy away in the more creative side. Throughout music history, many musical innovations started with an eccentric harmony that didn’t quite fit in with the general musical atmosphere at the time. Many groundbreaking composers in the past created music that was a little bit ahead of their time, and their adventurous exploration allowed future generations to keep pushing the definition of “good music.” For machines, their algorithms more or less base on existing data only, so this risk-taking mentality is difficult to acquire so far.
Technology has always been changing the music industry. When phonographs first appeared in the late 19th century, music publishers worried about the sales of sheets music. Sheets music did lose some popularity after phonographs, but records also appeared as a result and helped new genres of music to flourish nevertheless. [7] Artificial intelligence will surely change how musicians work with together on their music in the future, but the fast-paced and meticulous algorithms will help music professionals and educators to better create and spread their work. References: [1] Google Doodle, March 21, 2019 https://www.google.com/doodles/celebrating-johann-sebastian-bach [2] www.classicfm.com, “People are sharing their own versions of the Bach Google Doodle – and they’re really good,” March 22, 2019 https://www.classicfm.com/composers/bach/people-sharing-google-doodle-ai-game-compositions/?fbclid=IwAR3RnWwpSHkW77u9BZmVlG-xNevuVZGpoDU6iZhkH4nbcWwzYQg8ls-tMiI [3] Gino Brunner, Yuyi Wang, Roger Wattenhofer, Sumu Zhao, “Symbolic Music Genre Transfer with CycleGAN,” September 20, 2018 https://arxiv.org/pdf/1809.07575v1.pdf [4] Kin Wah Edward Lin, Balamurali B.T., Enyan Koh, Simon Lui, Dorien Herremans, “Singing Voice Separation Using a Deep Convolutional Neural Network Trained by Ideal Binary Mask and Cross Entropy,” December 4, 2018 https://arxiv.org/pdf/1812.01278v1.pdf [5] Jong Wook Kim , Justin Salamon, Peter Li, Juan Pablo Bello, “CREPE: A Convolutional Representation for Pitch Estimation,” February 17, 2018 https://arxiv.org/pdf/1802.06182v1.pdf [6] BBC News, “Robot conductor YuMi makes opera debut,” September 13, 2017 https://www.bbc.com/news/av/technology-41257534/robot-conductor-yumi-makes-opera-debut [7] Wade Roush, “Technology Is Upending How Music Is Made,” Scientific American, March 1, 2019 https://www.scientificamerican.com/article/technology-is-upending-how-music-is-made/
0 Comments
Leave a Reply. |
