Advancing Low-Resource Urdu Speech Emotion Recognition: Employing Deep Learning with Unidirectional and Bidirectional Gated Recurrent Units on Mel-Spectrogram Acoustic Features

This is the authors' abstract. We don't add key points for this paper.

In the growing field of artificial intelligence, identifying emotions in Urdu speech presents distinct challenges. This work adopts a novel approach using unidirectional (Uni-GRU) and bidirectional (Bi-GRU) gated recurrent units to examine the acoustic characteristics of the Mel-spectrogram for emotion classification. We propose a speech emotion recognition (SER) method based on these specific recurrent neural network architectures. Initially, the Uni-GRU model processed the Melspectrogram and effectively captured emotional patterns in URDU dataset samples, achieving 98.33% training accuracy, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7 5. 0 0 %</tex> validation accuracy, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">8 1. 2 5 %</tex> testing accuracy. Subsequently, the enhanced Bi-GRU model surpassed the Uni-GRU by more accurately identifying complex emotions, attaining a training accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">9 8. 7 5 %</tex>, validation accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7 8. 7 5 %</tex>, and testing accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">8 7. 5 0 %</tex>. Our evaluation on a representative Urdu corpus shows that GRU-based models substantially improve SER performance. This modeling strategy highlights the advantages of using Uni-GRU and Bi-GRU architectures with Mel-spectrogram features, offering insights into emotion recognition in low-resource languages. The findings demonstrate the potential of deep learning to support speech processing in underrepresented linguistic contexts and lay the groundwork for extending SER applications across a broader range of language environments.