Influence of Speakers' Emotional States on Voice Recognition Scores

摘要

The paper presents the voice recognition EER (Equal Error Rate) scores for speakers' basic emotional states. The database of Polish emotional speech used during the tests includes recordings of six acted emotional states (anger, sadness, happiness, fear, disgusts, surprise) and the neutral state of 13 amateur speakers (2118 utterances). The voice recognition procedure was proceeded with MFCC features and GMM classifiers. The EER scores distinctly depend on speakers' emotional states, even for a simulated database. The mean EER results tend to be only slightly less sensitive to an emotional state, even when using speech in various kinds of emotional arousal in a training set.