TY - GEN
T1 - Evaluating Speech Emotion Recognition Systems
T2 - 13th International Conference on E-Health and Bioengineering, EHB 2025
AU - Grosu, Mirela
AU - Mocanu, Bogdan
AU - Tapu, Ruxandra
AU - Datcu, Octaviana
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2026/1/1
Y1 - 2026/1/1
N2 - Speech Emotion Recognition (SER) plays a crucial role in affective computing, human-computer interaction, and psychological assessment. This paper presents a comparative analysis of traditional and transformer-based feature extraction techniques for SER. First, Mel-Frequency Cepstral Coefficients (MFCCs) and Mel-Spectrograms (Mel-Specs) are employed to train a three-layer Convolutional Neural Network (CNN) from scratch, assessing their effectiveness in classifying emotions. Subsequently, features extracted from a pretrained transformer model are evaluated using both Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) classifiers. Experimental results demonstrate that the MLP classifier applied to transformer-extracted features achieves the highest classification accuracy, reaching 89.94% on the RAVDESS and 84.75% on SAVEE. Finally, a detailed misclassification analysis reveals systematic confusion patterns, suggesting that SER models capture subtle affective cues beyond discrete emotion categories.
AB - Speech Emotion Recognition (SER) plays a crucial role in affective computing, human-computer interaction, and psychological assessment. This paper presents a comparative analysis of traditional and transformer-based feature extraction techniques for SER. First, Mel-Frequency Cepstral Coefficients (MFCCs) and Mel-Spectrograms (Mel-Specs) are employed to train a three-layer Convolutional Neural Network (CNN) from scratch, assessing their effectiveness in classifying emotions. Subsequently, features extracted from a pretrained transformer model are evaluated using both Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP) classifiers. Experimental results demonstrate that the MLP classifier applied to transformer-extracted features achieves the highest classification accuracy, reaching 89.94% on the RAVDESS and 84.75% on SAVEE. Finally, a detailed misclassification analysis reveals systematic confusion patterns, suggesting that SER models capture subtle affective cues beyond discrete emotion categories.
KW - Objective evaluation
KW - Speech emotion recognition
KW - Traditional low-level features
KW - Transformer model
UR - https://www.scopus.com/pages/publications/105044280091
U2 - 10.1007/978-3-032-24045-3_9
DO - 10.1007/978-3-032-24045-3_9
M3 - Conference contribution
AN - SCOPUS:105044280091
SN - 9783032240446
T3 - IFMBE Proceedings
SP - 81
EP - 88
BT - Advances in Digital Health and Medical Bioengineering 2 - Volume 2
A2 - Costin, Hariton-Nicolae
A2 - Magjarevic, Ratko
A2 - Petroiu, Gladiola Gabriela
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 13 November 2025 through 14 November 2025
ER -