An Artist Conducts a Generative Orchestra of Bells

Artist Budhaditya Chattopadhyay is known for exploring how technology can mediate a listener’s experience of sound through large-scale installation and live performance. 

Through attentive listening– and a collection of more than three terabytes of field recordings– Chattopadhyay transforms ephemeral experiences into sonic traces. 

His artworks have been exhibited, performed or presented in Transmediale, Berlin; ZKM Center for Art and Media, Karlsruhe; Donau Festival, Krems; Harvestworks, New York City; and Screen City Biennial, Stavanger, Norway, among others. 

Sound is my primary medium.

I'm deeply interested in pre-modern sounds, or sounds that existed before the intervention of modernist recording technologies. Like temple bells or wind chimes.

I work with field recordings and found objects. And I always carry a sound recorder wherever I go. 

In my work, I use creative coding, programming algorithms and generative systems to develop responsive compositions and sonic environments that address urgent issues, like the climate crisis, human intervention in the environment and ecology, migration, race, and decolonization.

I try to connect the disparate resonances across the borders and cultures through a practice committed to advocating for reciprocity and equality in contemporary societies.

For this project, I wanted to relocate  and re-listen to pre-modern sounds within our contemporary machine society and its automated environments. 

And I wanted to explore how machine learning might help me conduct an orchestra of such ritualistic sounds into an automated and self-regulating composition, thereby reconnecting them to the natural worlds from where they have been emerging.

Together with Google creative technologists Parag K. Mital and Holly Grimm, we explored ways to train a model on compositions using sequencing models. 

We trained a sequence model on two corpora (datasets) of individual bells recorded in different environments.

Each dataset was played in a different installation environment, and consisted of 50 stereo recordings, ranging from one to three seconds in length. 

These recordings were accompanied by a 10-minute “composition” wave file, which served as a ground truth sequence of sound for Google creative technologists to test the trained algorithm and see if our trained model performed well.