YSED: Yemeni speech emotion dataset

Arabic encompasses numerous dialects, including the Yemeni dialect, which is notably underrepresented in speech emotion recognition (SER) research. To support deep learning-based emotion recognition in Yemeni speech, there is a critical need for high-quality audio datasets that capture the emotional nuances specific to Yemeni Arabic. The limited availability of such resources has hindered research progress in this domain, particularly for Arabic and its dialects. In this paper, we introduce the Yemeni Speech Emotion Dataset (YSED), the first of its kind for Yemeni Arabic. YSED comprises audio recordings from 71 Yemeni volunteers (37 males and 34 females) aged between 15 and 45 years, expressing five emotions: anger, sadness, happiness, fear, and neutrality. Each participant recorded five sentences per emotion, with four repetitions for emotional states, and one for neutral, resulting in an initial 6035 sentence recordings. Recording duration ranged from 2 to 7 s using the simulated and induced approach. After evaluation by six judges, a total of 1432 recordings were validated, with an overall duration of 59 min and 10 s. The analysis achieved a Fleiss' Kappa coefficient of 0.9, indicating a high level of inter-annotator agreement. The YSED dataset is expected to be a valuable resource for researchers in emotion analysis and application development, enhancing the processing of Yemeni speech at an emotional level and informing services and social interactions. This work represents a significant step toward enriching data sources for understanding emotions in Yemeni Arabic and other languages and cultures. The dataset will be accessible at the following link: https://zenodo.org/records/15227219.