Cross-Lingual Transfert Learning for Speech Emotion Recognition

Cross-lingual Speech Emotion Recognition (SER), remains a challenge in real-world applications due to the availability of few labeled and many unlabeled features for datasets and a large shift distributions between the source and target domains. In this paper, we study the performance of cross-linguistic transfer learning through different approaches of feature learning namely features learning with Convolutional Neural Networks (CNN) and hybrid learning using hand-crafted 2D mel-spectrogram images and deep neural networks. The study results are evaluated by three datasets: Iemocap, Ravdess and EmoDB in both single and cross-corpus settings and compared using accuracy, loss, classification reports and confusion matrix evaluation metrics. The model achieved an accuracy of $82 \%$ on Ravdess, $83 \%$ on Emodb and $78 \%$ on IEMOCAP.