I am working on developing a biologically inspired feature-based model for speech categorization and comparing its performance with modern machine learning models.
I have also been modeling neural mechanisms underlying robust auditory categorization of complex signals. Specifically, I have been focusing on how bottom-up and top-down mechanisms aid in noise-invariant and reverberation-invariant processing of vocalizations signals, and relating the model performance with animal behavioral performance to elucidate mechanisms that best explain behavior.
My research focus was to quantify speech coding fidelity in the normal and impaired auditory systems including models of noise-induced and Carboplatin-induced hearing loss. I performed invasive neurophysiological experiments to record from auditory-nerve fibers and noninvasive electrophysiological experiments to record frequency-following responses using natural speech as the stimulus.
I developed spectrotemporal tools that improve the translational aspects of animal models by linking invasive single-unit responses and noninvasive evoked responses.
I investigated the degradation of tonotopic coding of speech following hearing loss in both single-unit and electrophysiological data. These insights led to the development of a noninvasive diagnostic measure to test for distorted tonotopy in humans.
I also studied the neural correlates of perceptual deficits experienced by patients with hearing loss in noisy environments. In particular, I quantified the strength of perceptually important speech features in spike-train data when speech is presented in the backgrounds of steady-state and fluctuating noise.
During my stay at DTU, I worked on extending the multi-resolution speech Envelope Power Spectrum Model from the acoustic domain to the neural domain using collected and simulated spike train data. This extension will allow estimating speech intelligibility under various degraded conditions using neural spike trains collected from animal models of sensorineural hearing loss.
During my undergraduate thesis work, I worked with simulated and experimental neural data from the dorsal cochlear nucleus in response to high dimensional non-Gaussian stimulus space (random-spectral shaped stimulus).
I implemented a gradient descent algorithm to find maximally informative spectral dimensions for individual neurons using a spike-count model.
Using machine learning and acoustic-to-articulatory inversion methods, I created a GUI to generate spatially-rich midsagittal view of articulatory trajectories at high temporal resolution from recorded speech on the fly.