AN IMPROVED DEEP LEARNING TECHNIQUE FOR SPEECH EMOTION RECOGNITION AMONG HAUSA SPEAKERS

This research addressed the challenge of recognizing emotions from speech by developing a deep learning-based speech-emotion recognition (SER) system. A key focus of the study is the creation of a new Hausa emotional speech dataset, aimed at addressing the linguistic and cultural imbalance in existing SER datasets, which predominantly feature Western languages. This study captured four emotions: happy, sad, angry, and surprise among native Hausa speakers. The self-captured dataset was recorded in an environment that is devoid of noise to ensure high quality and uniformity in the audio data. A public dataset; RAVDESS was used for benchmarking the proposed technique. CNN and Bi-Directional Long Short-Term Memory (BiLSTM) architectures were combined and used as proposed model for the SER experiment. The developed CNN architecture helped in extracting spatial features, while the BiLSTM without the attention mechanism captured temporal dependencies from the audio data. The approach reduced time complexity and improved performance to 100% and 96% recognition accuracies against 94% and 90% of the benchmark model for the local and benchmark datasets respectively. The results demonstrate the proposed approach's robustness to generalize across linguistic contexts.