Author
Bozkurt, E., Erzin, E., Eroğlu Erdem, Ç., Erdem, Tanju
Publication Date
2010
Publication Place
-
IEEE
Subject
Gaussian processes, Emotion recognition, Feature extraction, Hidden Markov models, Pattern classification, Spectral analysis, Speech recognition, Unsupervised learning
Type
Document
Language
Turkish
Digital
Yes
Manuscript
No
Library
Özyeğin University
Library Asset ID
978-1-4244-9672-3
Record ID
45a3a710-d49f-42d5-ad4d-a51fd28a6a6c
Library Location
Computer Science
Date
2010
Notes
Due to copyright restrictions, the access to the full text of this article is only available via subscription.
Sample Text
In this article, we evaluate the results of the INTERSPEECH 2009 Emotion Recognition Competition. The problem presented by the competition is the most accurate separation of natural and emotion-rich FAU Aibo speech recordings into five and two emotion classes. To solve this problem, we examine many correlated, spectral and SMM-based (hidden Markov model) features with Gaussian Component Model (GBM) classifiers. While the spectral features include the Mel frequency cepstral coefficients (MFKK), the true spectral frequency (DSF) coefficients and their derivatives, the hymen features consist of pitch, first derivative of pitch and energy. We obtain SMM features, which describe the change of all related features over time, with unguided trained SMM structures. We also examine the data fusion of different features and the decision fusion of different classifiers to improve emotion recognition results from speech. Our two-stage decision fusion method achieved a success rate of 41.59% and 67.90% for five- and two-class problems, respectively, and ranked 2nd and 4th among all competition results.
DOI
10.1109/SIU.2010.5649919
