March 2017 codeaudio

A large-scale dataset of sounds useful for machine learning produced by Google. I designed the website and helped with the research.

Quick links:

A screenshot of the website, showing the text 'Evaluation segments for Violin, fiddle,' and YouTube thumbnails corresponding to video segments that can be played.

An open source research project I worked on for over a year at Google. My main tasks included designing and implementing the frontend human annotation interface, helping analyze data and construct the ontology, implementing the research website, and assisting with the technical write-up.

A histogram showing the number of examples for each audio class. 'Music' has a little over a million at the top, and 'toothbrush' only has 127. A graph depicting an ontology of sound classes. The top nodes of the graph read 'Human sounds,' 'Animal,' 'Music,' 'Source-ambiguous sounds', 'Sounds of things,' 'Natural sounds,' and 'Channel, environment and background.'