Speech emotion recognition in noisy environments using adaptive comb filtering and feature vector classification

Abstract

Speech processing technology has been continuously advanced as an essential field of artificial intelligence for the last few decades. Researchers are investigating various approaches to make more intelligent and user-friendly applications, which assist and understand humans. Especially, service robot industry remains critical challenges to advance the human-robot interaction and they search for the solutions in voice interface. The most representative technology to enable intelligent machines to comprehend users` emotional state and interact with them is Speech Emotion Recognition (SER), which automatically identifies users` feelings and conditions from their spoken speech. Although many researchers have applied various technical approaches developed for speech recognition and speaker identification to SER system, they should take into account two critical issues. The first one is noisy environments. Many applicable devices of SER such as mobile devices and service robots are inevitably exposed to various background noises. As studied on speech recognition, noise contaminated speech may significantly degrade the recognition performance, and this drawback should be handled in SER equivalently. The next issue is the domain-oriented ambiguity: acoustically similar characteristics between emotions and variable speaker characteristics due to different user speaking styles. Each of these characteristics may cause a substantial amount of overlap between emotion models in feature vector space, thus decreasing SER accuracy. This dissertation aims at preserving the SER performance from each issue mentioned above. To address noisy environmental issue, this dissertation proposes an efficient front-end of SER system. The proposed front-end is based on adaptive comb filtering. Whereas conventional adaptive comb filtering reduces noise components remarkably, it is rarely effective in reducing severe noises. Furthermore, due to the uniformly distributed frequency response...