Επαλήθευση ομιλητή με χρήση συναισθήματος

Κατσίρος, Δημήτριος; Katsiros, Dimitris

dc.contributor.author	Κατσίρος, Δημήτριος	el
dc.contributor.author	Katsiros, Dimitris	en
dc.date.accessioned	2022-01-20T08:52:57Z
dc.date.available	2022-01-20T08:52:57Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/54373
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.22071
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc/3.0/gr/	*
dc.subject	Speaker verification (SV)	en
dc.subject	Επαλήθευση ομιλητή	el
dc.subject	Text-Independent speaker verification (TISV)	en
dc.subject	Speech E- motion recognition (SER)	en
dc.subject	Emotion driven speaker verification	en
dc.subject	Επαλήθευση ομιλητή ανεξαρτήτως κειμένου	el
dc.subject	Αναγνώριση συναι- σθημάτων oμιλίας	el
dc.subject	Επαλήθευση ομιλητή με χρήση συναισθήματος	el
dc.title	Επαλήθευση ομιλητή με χρήση συναισθήματος	el
dc.title	Emotion driven speaker verification	en
heal.type	bachelorThesis
heal.classification	Deep learning	en
heal.classification	Machine learning	en
heal.classification	Speaker verification	en
heal.classification	Speech emotion recognition	en
heal.language	el
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2021-07-19
heal.abstract	Speaker Verification (SV) enables the authentication of a claimed identity from measurements on a voice signal. Emotion as a natural and often involuntary encoder of voice, has the mechanisms responsible for vocal modulation. Despite the attention that the field has gained over the years, little effort has been made in order to identify the relations between these two subjects. Αlthough seemingly far, emotional content could have a huge impact on speaker discrimination. In this thesis, we investigate the correlation between speaker verification and speech emotion recognition. First of all, we create various emotional evaluation sets, each one aiming to track differently the effect of emotion on the speaker verification task. In an attempt to decrease or even eliminate the effect we try to transfer emotional knowledge to our task. For this purpose, we implement four different architectures, each one of them, handling emotional information in a different manner. Then we examine our models’ performance on the emotional evaluation sets. Our results suggest that emotional information has a crucial role on speaker verification. Even on low intensity, emotion on both on enrollment and verification can significantly degrade a system’s performance. On addition, emotions on strong intensity, seem to escalate the effect and ensue in poor results. Among the seven emotions examined, we find that, anger and fear were these having the most remarkable impact. In an endeavor to address the aforementioned issues we examine the performance of our emotion-aware architectures. Our results indicate that by applying classic fine tuning techniques, we are able provide emotion robust models and at the same time perform much better on the speaker verification task. Last but not least, we test our hypothesis on provi ding same-emotion utterances on evaluation phase and we observe a relative improvement around 20%, irrespective of emotional pre-training. Overall, we can capture a strong relation between speaker discrimination and emotional content. We contend that controlling emotional content is necessary for a model’s robustness, especially for real life scenarios, where emotion is present. Ultimately, we can reduce the effect and improve our models performance by applying traditional transfer learning techniques from speech emotion recognition to speaker verification.	en
heal.advisorName	Ποταμιάνος, Αλέξανδρος	el
heal.advisorName	Γιαννακόπουλος, Θεόδωρος	el
heal.committeeMemberName	Τζαφέστας, Κωνσταντίνος	el
heal.committeeMemberName	Ποταμιάνος, Αλέξανδρος	el
heal.committeeMemberName	Γιαννακόπουλος, Θεόδωρος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών. Τομέας Σημάτων, Ελέγχου και Ρομποτικής	el
heal.academicPublisherID	ntua
heal.numberOfPages	87 σ.	el
heal.fullTextAvailability	false