Usually, people talk neutrally in environments where there are no abnormaltalking conditions such as stress and emotion. Other emotional conditions thatmight affect people talking tone like happiness, anger, and sadness. Suchemotions are directly affected by the patient health status. In neutral talkingenvironments, speakers can be easily verified, however, in emotional talkingenvironments, speakers cannot be easily verified as in neutral talking ones.Consequently, speaker verification systems do not perform well in emotionaltalking environments as they do in neutral talking environments. In this work,a two-stage approach has been employed and evaluated to improve speakerverification performance in emotional talking environments. This approachemploys speaker emotion cues (text-independent and emotion-dependent speakerverification problem) based on both Hidden Markov Models (HMMs) andSuprasegmental Hidden Markov Models (SPHMMs) as classifiers. The approach iscomprised of two cascaded stages that combines and integrates emotionrecognizer and speaker recognizer into one recognizer. The architecture hasbeen tested on two different and separate emotional speech databases: ourcollected database and Emotional Prosody Speech and Transcripts database. Theresults of this work show that the proposed approach gives promising resultswith a significant improvement over previous studies and other approaches suchas emotion-independent speaker verification approach and emotion-dependentspeaker verification approach based completely on HMMs.
展开▼