ASR is used heavily in eyes-busy or hands-busy situations and often time the user may be speaking over noise. My peers and I are particularly interested in how music effects ASR decoding. We use several music datasets of varied genre or broken-down instrumentation to allow us to perform in-depth anaysis of how different aspects of music influences speech recognizer’s performance. We then train a new model from what we have learned to see if we could improve the original model’s performance.

EEGs are extremely interesting to the signal processing world as the signals from them are high dimensional and localized spatially, spectrally, and temporally. These signals are extremely low amplitude and artifacts appear in the measured signal due to normal human processes like blinking, breathing, and moving; or noise can appear due to the external factors such as line noise or sound in the room. My peers and I implement Viola’s ICA CorrMap procedure to denoise EEGs, but rather than retreiving EEG artifacts in-situ, we use Cho’s EEG dataset which separately records subjects in both rest states and performing various ‘noisy’ actions prior to a motor imagery trial. We use this data to try and denoise signals and try to devise a way to create a generalized artifact template that can be transferred from user to user.

I developed image processing pipelines for autonomous vehicles (AVs) and architected an extensible simulator for AVs using Microsoft AirSim and Unreal Engine.