Srinath Tankasala
stankasa@utexas.edu
In this project I trained the RL locomotion policy on a Laikago quadruped robot in simulation
In this internship I explored the ability of different LLMs to re-write transcripted speech to correct for different types of errors. Exciting work coming ahead in the next few months on LLM generalization capablity to different tasks (such as ASR correction) with different input modalities.
This internship was my first venture into improving Automated speech recognition (ASR) using graph based methods. Specifically, we showed for the first time that information from the utterance audio can be shared across each utterances using graph based label propagation. This information sharing was used to re-rank the N-best hypotheses leading to fairer and improved WER and SER performance. I had the wonderful opportunity of working with amazing collaborators leading to a top 3% paper at ICASSP 2023!
I had the wonderful opportunity of working on SLAM optimization during my internship. I worked on improving the 3D reconstruction quality of different structures by imposing geometric constraints.
I worked with Robotics and Controls teams at MathWorks. I worked on better integrating ROS and MATLAB as part of the ROS toolbox.