A Multilingual Framework of CNN and Bi-LSTM for Emotion Classification

Speech emotion recognition is a crucial task for developing human-computer based interfaces. However, many significant studies have focused on extracting low-level features for classification. In this paper, we propose a language-independent, deep learning-based framework for speech emotion classification. We extract the MFCC based features along with deep high-level features by developing a unique combination of 1D CNN and Bi-LSTM units. The proposed framework can extract the local information from the signals with the help of CNN, and the Bi-LSTM layer helps in modeling the long-term contextual dependencies of the signal. We validate our proposed framework on two multilingual datasets, namely: EmoDB for German language and RAVDESS for the English language. Experimental results indicate that EMODB and RAVDESS datasets achieve an average accuracy of 94% and 73%, respectively. We also compare our results with state-of-the-art and alternative methods to validate our framework and measure the performance of the algorithm in terms of precision, recall, F1 score, and accuracy.