High Performance Speech Recognition Using Consistency Modeling

Vassilios Digalakis, Hy Murveit, Mitch Weintraub

Abstract: The primary goal of this project is to develop acoustic modeling techniques that advance the state-of-the-art in speech recognition, focusing on those techniques that relax the hidden Markov model's improper independence assumptions. Such techniques should both improve robustness to systematic variations such as microphone, channel, and speaker, by conditioning state's acoustic output distributions on long-term measurements, as well as improve general acoustic calibration by removing improper short-term (e. g. frame to frame) independence assumptions.