Comparative Analysis of Advanced ML Techniques for Urdu Speech Emotion Recognition: An Acoustic Study in Low-Resource Languages*

This paper presents a comprehensive comparative analysis of machine learning techniques for emotion recognition in Urdu speech, a domain that presents unique challenges associated with low-resource languages. We employ Mel- frequency cepstral coefficients to the rich SEMOUR PLUS dataset and explore the key models including support vector machines, k-nearest neighbors, long short-term memory networks, convolutional neural networks, dense neural network, few-shot learning, and Siamese network. Each model’s performance and adaptability in Urdu speech emotion recognition are rigorously assessed. The research not only underscores the potential of machine learning in emotion recognition but also spotlights the intricacies of deploying these techniques for languages with limited resources. As a prelude to future avenues of exploration, we delve into the potential of an integrated audio-centric emotion recognition system. Our findings serve as a foundational stone in the expanding realm of Urdu speech emotion recognition and offer insights for ensuing studies in similar languages.