Αναγνώριση Συναισθήματος μέσω της Φωνής

Χάσπαρη, Θεοδώρα; Chaspari, Theodora

dc.contributor.author	Χάσπαρη, Θεοδώρα
dc.contributor.author	Chaspari, Theodora
dc.date.accessioned	2025-06-17T06:57:04Z
dc.date.available	2025-06-17T06:57:04Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/62075
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.29771
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	συναίσθημα στον λόγο	el
dc.subject	Δυναμικά χαρακτηριστικά	el
dc.subject	Πίσσα	el
dc.subject	Emotion Recognition	en
dc.subject	Pitch	en
dc.subject	Formants	en
dc.subject	AM-FM demodulation features	en
dc.subject	GMMs	en
dc.title	Αναγνώριση Συναισθήματος μέσω της Φωνής	el
dc.contributor.department	Τομέας σημάτων ελέγχου και ρομποτικής	el
heal.type	bachelorThesis
heal.classification	Συναισθηματική αναγνώριση	el
heal.language	el
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2010-09-01
heal.abstract	Emotion Recognition is a part of aﬀective computing, which focuses on facilitating com-munication between humans and computers. In this diploma thesis, we examine emotion recognition based on speech. More speciﬁcally, basic features of emotion recognition, such as pitch, formants and utterance duration, are studied analytically. Besides these, it is supported that AM-FM modulation features can distinguish the ﬁne variations of emotions in speech. Instant amplitude and frequency are computed through the Energy Separation Algorithm (ESA) based on the Teager Energy Operator (TEO). Statistical moments of them are used as features. These features are strongly smoothed with median ﬁltering in order to remove the redundant information and keep only the essential information needed for emotion recognition. Experiments are conducted in two databases: the Berlin Database and the Aiginiteio Hospital Database of Emotional Speech, which include seven and ﬁve classes of basic emotions respectively. Classiﬁcation is done with K-means algorithm, GMMs based on expectation maximization and dynamically modiﬁed GMMs. Results vary from 30%to 90%. The most powerful features, which produce the best results, seem to be the TEO-Autocorrelation-Envelope, the Area of Instant Amplitude and the Weighted Mean of Instant Frequency.	en
heal.sponsor	ΕΜΠ	el
heal.advisorName	Μαραγκός, Πέτρος Α.
heal.committeeMemberName	Μαραγκός, Πέτρος Α.
heal.committeeMemberName	Πρωτόπαπας, Αθανάσιος
heal.committeeMemberName	Ποταμιάνος, Γεράσιμος
heal.academicPublisher	Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	177 σ.
heal.fullTextAvailability	false