Abstract
To make music composition more approachable, we designed the first AI-poweredGoogle Doodle, the Bach Doodle, where users can create their own melody andhave it harmonized by a machine learning model Coconet (Huang et al., 2017) inthe style of Bach. For users to input melodies, we designed a simplifiedsheet-music based interface. To support an interactive experience at scale, were-implemented Coconet in TensorFlow.js (Smilkov et al., 2019) to run in thebrowser and reduced its runtime from 40s to 2s by adopting dilated depth-wiseseparable convolutions and fusing operations. We also reduced the modeldownload size to approximately 400KB through post-training weight quantization.We calibrated a speed test based on partial model evaluation time to determineif the harmonization request should be performed locally or sent to remote TPUservers. In three days, people spent 350 years worth of time playing with theBach Doodle, and Coconet received more than 55 million queries. Users couldchoose to rate their compositions and contribute them to a public dataset,which we are releasing with this paper. We hope that the community finds thisdataset useful for applications ranging from ethnomusicological studies, tomusic education, to improving machine learning models.