Emotion recognition from speech is an important and challenging task, with applications in various domains of human-computer interaction. In this work, we propose a method that transforms speech signals into the frequency domain to extract meaningful spectral features, which are then classified using four learning models: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network (CNN). Our experiments utilize four standard datasets alongside a custom dataset, consisting of audio recordings representing three common emotional states: anger, happiness, and neutrality. Evaluation based on Accuracy, Recall, Precision, and F1-Score demonstrates that the CNN model significantly outperforms the others, achieving an accuracy of 100% on unseen test sets. These results confirm the strong potential of deep learning approaches for reliable and accurate speech-based emotion recognition.