Using CNN and Transformer Model for Unimodal Speech Emotion Recognition on MELD and IEMOCAP

Emotion Recognition is one field that is taking the world by storm in this current age. Multimodal emotion recognition has shown promising results however, previous studies shows that recognition using speech is a field has not performed that greatly. Previous research used popular methods such as CNN, RNN, and other machine learning methods has been used to achieved good results with the help of other modality, where if speech were to be used alone the results won't be as good. A factor that contributes to the result of the speech modality not performing as well is the fact of the dataset and the way it was processed. Therefore, this research looks into what makes speech one of the lowest performing modalities by preprocessing the dataset using a different method than previous researches as well as using a CNN model paired with a Transformer model. Results achieved from this research showed an improvement when using IEMOCAP by 8.09% and MELD by 1.06%. A comparison results of using the pre-processed data with previous studies model was also done to compare the results. In conclusion, this research has explored a different preprocessing method and deep learning model that has achieved results better than previous works with the same dataset.