Project AEI

Project AEI, backed by UC Berkeley’s Greater Good Science Center, is an effort to detect and deter police brutality using love-streamed data from officer body cameras. We engineered this problem as an emotion classification problem, using the streamed audio from the body cams to predict the emotional state of the relevant officer. We hypothesized that when certain emotions such as anger were detected from our data, the likelihood of violence was higher than other emotional states we could classify the audio data into. Detecting such information in real time could lead to preventative action to avoid violence.

I was a researcher and software engineer for the project. On the software engineering side, I created a website for the project and an interactive demo for our models. On the research side, I increased the classification accuracy of our model by introducing greater dataset diversity, optimizing data pre-processing, and trying out new model architectures and improvements. Our data pipeline consisted of transforming audio data into mel spectrograms before running the processed data through a CNN trained for our specific downstream classification task.