When Queen guitarist Brian May was creating the thunderous “Thump thump clap” sound that would become the introduction to their megahit We Will Rock You, his background in physics helped him immensely. May used a series of individual overdubs to build the grand and reverberating sound on the stamping beats. “Later, people designed a machine to do it,” May quipped regarding the manual layering process. Indeed, new technologies have fundamentally changed the music industry since the 80s. And a new machine is now creating a buzz in the music community.
Magenta Studio is a Google Brain project “exploring the role of machine learning as a tool in the creative process.” The Google Brain team created the open-source music-making package using machine learning models. The suite includes four tools: Continue, Generate, Interpolate, and GrooVAE. Musicians can use the models on their MIDI (Musical Instrument Digital Interface) clips. Currently, the suite of tools is available as a standalone Electron application and as a plugin for the software music sequencer and digital audio workstation Ableton Live, which was initially designed as an instrument for live performances but is now also being widely used for music composing, recording, arranging, mixing and mastering.
“In my opinion, the most interesting feature of Magenta is GrooVAE” says Toronto-based music producer Ryan McDougall. “Currently, other apps can only humanize the sound by putting in random variations. Magenta Studio seems different by using neural network training.”
Although advanced technologies have improved music production in terms of efficiency and accuracy, digital processes can lack the slight but familiar “dropped beat” feel and other natural imperfections that are generated by human players. GrooVAE can make a drum clip sound and “feel” like a human drummer’s performance. Magenta researchers quantized 15 hours of real drummers performing on MIDI drum kits, then trained a neural network to predict the beats. They used a predictive method similar to Google’s automatic language translation system, with drum beats and patterns taking the place of words and phrases. Google Brain researcher and guitarist Jesse Engel tweeted that GrooVAE’s beat is so humanlike it “almost makes me shudder.”
Magenta Studio can also be a useful tool for musicians looking to break through creative blocks. The Continue tool uses the predictive power of recurrent neural networks (RNN) to generate up to 32 measures of additional music following the original drum beat or melody inputs. Moreover the process does not only perform simple predictions of the melody, it also creates variations on themes.
For wholly blocked musicians who are unsure even how to start, Magenta’s Generate tool will produce a four-bar phrase with no input provided. This is done by a Variational Autoencoder (VAE) trained on millions of melodies and rhythms which has learned a summarized representation of musical qualities from the data.
Merging different musical ideas is another challenge in music composition and production, and Magenta’s Interpolate tool assists musicians in this regard. Taking two drum beats or two melodies as input, Interpolate creates clips which seamlessly integrate qualities from both inputs. The VAE also plays an important role here, enabling Interpolate to specify musical patterns by mapping connections between the inputs and eventually clustering them into a compressed MIDI profile.
Magenta Studio is not the first AI to tackle music making. American Singer Taryn Southern used Magenta, virtual composer AIVA, IBM’s Watson Beat, and AI composer and API Amper on her 2018 album I Am AI. Amper is another music software company using AI to create music in a range of genres and moods.
An increasing number of AI-powered startups are making noise across the music industry, while their new AI music generating techniques also have commercial applications in film and television background music, soundtracks, video games, etc. As one of the pilot applications from Google’s Magenta project, Magenta Studio represents the latest effort from tech giants to explore AI’s potential in the creative arts.
Journalist: Fangyu Cai | Editor: Michael Sarazen