Аннотация
Speech Emotion Recognition (SER) plays a critical role in affective computing and human-computer interaction. This study introduces a novel and fully automated SER framework that integrates Convolutional Neural Networks (CNNs) with advanced feature selection techniques to enhance classification accuracy while maintaining computational efficiency. Leveraging the librosa toolkit, five spectral feature types—including Mel-frequency Cepstral Coefficients (MFCCs), chroma, mel-spectrogram, spectral contrast, and tonnetz—were extracted, resulting in a 193-dimensional feature space. The Fisher score algorithm, combined with Recursive Feature Elimination (RFE), was used to identify the most discriminative features. Additionally, a gender-specific preprocessing pipeline and recognition have been integrated to mitigate bias and improve model robustness. The proposed model was evaluated on three benchmark datasets: RAVDESS, EMO-DB, and EMOVO, covering eight distinct emotional states. The results demonstrate strong generalization across languages and speaker identities, validated through both conventional and leave-one-out cross-validation settings. These findings underscore the effectiveness of combining deep learning with feature selection for scalable, accurate, and interpretable speech emotion recognition.