همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Persian Speech Emotion Recognition Using Multi Scale SincNet and cGAN-based Ensemble Learning

نویسندگان :

Ali Hekmat¹ Yasser Shekofteh²

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی

کلمات کلیدی :

Speech Emotion Recognition،SincNet،cGAN،Ensemble Learning،Interpretability،ShEMO

چکیده :

Abstract—Speech Emotion Recognition (SER), particularly for the Persian language, presents significant challenges due to data scarcity and emotional variability among speakers. This research proposes a novel multi-stage framework to enhance the accuracy, interpretability, and generalizability of Persian SER models. The proposed architecture consists of three primary stages: (1) feature extraction using a multi-scale interpretable SincNet, (2) speech and feature map reconstruction via conditional Generative Adversarial Networks (cGANs), and (3) a final classification decision based on a One-vs-All architecture with ensemble learning using hard and soft voting. To ensure a fair and robust evaluation, experiments were conducted on the ShEMO dataset using a speaker-independent configuration, which is critical for real-world applicability. The results demonstrate a significant improvement over baseline methods across all key evaluation metrics. The proposed method achieved a Weighted Accuracy of 86.35%, a 1.56% increase over the best-performing baseline, HuBERT, which reached 84.79%. Furthermore, our approach achieved a Weighted F1 score of 86.36%, representing a 2.32% improvement over the HuBERT baseline's score of 84.04%.