<p>Speech Emotion Recognition (SER) is an important field in human–computer interaction and remains a challenging task. The biggest challenges in SER analysis include differences in language, accent, and recording conditions. Therefore, in this study, three different datasets were used, and spectrograms were employed due to their effectiveness in representing speech signals. In real-world scenarios, noise is often mixed with speech. To better simulate such conditions, data augmentation was applied by adding real-world environmental noise, white Gaussian noise, and time shifting to the audio signals. Both the original and augmented audio files were then converted into spectrograms using the Short-Time Fourier Transform (STFT). To analyze emotions from these spectrograms, a four-layer CNN model was designed and evaluated using the RAVDESS, TESS, and SAVEE datasets. Rather than increasing model depth, a compact architecture was preferred to examine the effect of kernel size and filter progression under limited data conditions.The model achieved 100% accuracy on the TESS dataset, 99.51% on RAVDESS, and 95.57% on SAVEE. However, the high performance on TESS should be interpreted with caution due to its controlled recording conditions and limited speaker diversity. The results indicate that an appropriately configured low-depth CNN can provide competitive performance without substantially increasing model complexity.</p>