Can machines learn and make music? The question always sparks aesthetic and philosophical arguments, but Google’s Magenta project is examining whether machine learning can help “to create compelling art and music,” according to its website.
Some representatives from Magenta were at The Cage at the American Tobacco Campus Saturday demonstrating some of the sound possibilities of their research during the AI [Artificial Intelligence] Jam Session and NSynth Demo during the third day of Moogfest 2017.
Colin Raffel of Magenta told some festivalgoers that the hardware they were getting to touch and explore had been “taught” the sounds of different instruments. The goal is to provide “creative tools for musicians and artists,” Raffel said. Raffel was demonstrating the work of the program called Melody RNN, and one advantage of the program having “learned” certain sounds is that it will correct “and even embellish” the notes that users who may not have great musical skill may produce.
Jesse Engel of Magenta was demonstrating how the NSynth program had learned a set of sounds. He took a sample of a violin and a flute, and showed how those combined sounds could be manipulated to create other sounds and timbres. He also manipulated the sound of a cat’s meow and a flute. When we hear these sounds in audio, we still hear distince sounds, Engel said. NSynth “is creating a single sound out of two sounds,” he said. With machine learning, the computer also “makes a next prediction [about] what is the next sound,” he said.
Several visitors had their notebooks (paper, not digital) out as Engel explained Magenta’s open source software and how it can be downloaded. One of them was Gina Likins, who works for Red Hat software in Raleigh. Like Red Hat’s software, Magenta is open source, and what Magenta is doing with machine learning reflects the community aspect of open source development, Likins said. Machine learning has become more accurate because of many different programmers sharing their work, she said.
Likins teaches workshops on how to teach open source software. She envisions a way to allow students in music schools to input sounds and phrases into Magenta’s programming “and increase [the machine’s] musical literacy by magnitudes,” Likins said. “How can you grow the number of people who are active contributors to the project?” she asked. “So many music schools now are trying to figure our how this works into their curriculum.”
Tim Draegen, a software engineer from Brevard, said he was fascinated by machine learning because of its ability to mix different instruments in different ways. “I’m here [at Moogfest] because I don’t know much about the synth world, so this is my introduction,” Draegen said.
For more information about Magenta, visit magenta.tensorflow.org/nsynth-instrument