Towards Sustainable Development in Speech Recognition: Enhancing Emotional Understanding in Bilingual Societies

Objective: The aim of this work was a comparative analysis of the effectiveness of emotion recognition models based on prosodic, spectral, and combined features for Armenian and Kabardian languages. Method: An experimental corpus of speech utterances in Armenian and Kabardian languages was collected, containing samples of emotional (anger, joy) and neutral speech. From these speech signals, prosodic features (fundamental frequency, intensity, duration) and spectral features (MFCC) were extracted. Results and Discussion: Three machine learning models were trained: a model based on prosodic features, a model based on spectral features, and a model based on a combination of features. The effectiveness of the models was assessed using precision and recall metrics for emotion recognition. Research Implications: The results showed that the model based on a combination of prosodic and spectral features demonstrates the best precision and recall rates for both anger and joy emotions in both languages.