Audio-Based Emotion Analysis Using Deep Learning Techniques

This paper uses audio parameters such as MFCCs, zero-crossing rate, chroma features, RMS values, and Mel spectrograms to provide a novel approach to machine learning for speech emotion recognition. Heterogeneous audio recordings from datasets RAVDESS, CREMA-D, TESS, and SAVEE are used for the experiment. Data augmentation of the dataset was done by performing pitch modification, stretching, shifting, and addition of noise in order to improve model robustness. CNN, being competitive with regard to emotion categorization and sequential data analysis, is being explored for possible application to sentiment analysis, mental health monitoring, or even human-computer interaction; the work, thus, makes it possible for advancing a scalable, intuitive, yet accurate system in speech emotion recognition.