Procuring MFCCs from Crema-D Dataset for Sentiment Analysis using Deep Learning Models with Hyperparameter Tuning

Sentiment analysis is indispensable for scrutinizing the convoluted characteristics of the human mind. In this paper, an attempt is made to perform a comprehensive analysis of sentiment in the CREMA-D audio dataset by relying on extracting MFCC features from the dataset. Standardization of the dataset is carried out owing to the unequal length of each file. MFCC feature extraction on the standardized dataset leads to the harvesting of attributes such as pitch, timbre, formants etc. which are crucial for the identification of each unique audio file and their corresponding sentiment. A Deep Learning based framework is proposed to carry out this examination on the dataset by executing Convoluted Neural Networks (CNN), Recurrent Neural Networks-Long Short-Term Memories (RNN-LSTM), and Recurrent Neural Networks-Transformers (RNN-Transformers) algorithms. Six classes have been employed for this analysis, which are happy, sad, fear, disgust, angry and neutral. Maximum accuracies of 56.7% and 54.0% have been obtained from CNN and RNN models respectively.