Adaptive & discriminative speech modeling to cope with temporal changes of environments

Abstract

This research focus is motivated by the general issue of robust acoustic modeling of speech for achieving superior and reliable performance across various application environments. In real environments, characteristics of a speech signal vary considerably, due to changes of contents, speakers, and ambience. There is thus an acute need in the methodology for the construction of robust and high-accuracy speech models that adaptively respond to these changes of the environment. In particular, the approach that is being taken in this collaboration is based on discriminative modeling. We aim to apply our studies and new techniques to the task of speaker diarization and acoustic event detection in addition to speech recognition.