Paul Hager
MIT EECS — Lincoln Labs Undergraduate Research and Innovation Scholar
Investigation of Deep Neural Networks in Speech Processing
2015–2016
James Glass
The recent resurgence of deep neural networks has sparked questions in classical statistical inference methods. A particular inference problem is dialect recognition, which alongside international communication has become increasingly important in recent times. My work utilizes semi-supervised learning to harness large amounts of unlabeled data and explores the strengths of novel network architectures including long-short term memory networks to improve accuracy in dialect recognition. Using standard techniques I also investigate methods to train these deep network architectures with a limited amount of data. Throughout my project I will extend DNN toolkits and kaldi recipes for studying speech recognition problems in support of the Spoken Language Systems Groups work.
I participated in the Jelinek Speech and Language Technology Workshop with a team developing an automatic speech recognition system trained exclusively from transcripts written by non-native speakers. My research interests lie in how to develop systems machines can utilize to ease their interactions with humans. I am excited to further the goal of allowing humans and machines to seamlessly communicate through speech.