<![CDATA[Speech Emotion Recognition (SER) is a crucial area in human-computer interaction, where selecting effective feature representations remains a key challenge. This study proposes an SER system using Log-Mel Spectrogram images and a 2D Convolutional Neural Network (CNN). The Toronto Emotional Speech Set (TESS) dataset, comprising 2,800 audio samples across 7 balanced emotion classes (400 samples each), was employed. Each audio file was converted into a 128×128-pixel Log-Mel Spectrogram image. The CNN architecture consists of three convolutional blocks (32, 64, and 128 filters) with Batch Normalization and Max Pooling, followed by a Fully Connected layer (128 neurons), Dropout (rate 0.5), and a 7-class Softmax output layer. The model was trained using the Adam optimizer, with early stopping triggered at epoch 22. Experimental results demonstrate that the model achieves 100% classification accuracy on 560 test samples, yielding perfect precision, recall, and F1-score (1.00) for all classes. This research confirms that visual spectrogram representations combined with CNNs can automatically and effectively extract emotional features without manual feature engineering.]]>
Copyrights © 2026
