Speech Emotion Recognition (SER) plays a vital role in enhancing human-computer interaction by enabling machines to interpret and respond to human emotions. This study focuses on SER using the RAVDESS dataset, emphasizing speech-only modalities. A comprehensive set of audio features including MFCCs, chroma, spectral contrast, tonnetz, and wavelet transforms is extracted, and the performance of four deep learning models—CNN, LSTM, BiLSTM, and CNN-LSTM—is evaluated. Among them, CNN achieves the highest accuracy (68.37%), with strong F1-scores across several emotion classes. The results underscore the effectiveness of spatial feature extraction in emotion classification and suggest further enhancements using attention mechanisms and transformer-based models.