Speech Emotion Recognition using CNN

Abstract Speech Emotion Recognition (SER) is a growing area in affective computing that aims to detect and understand human emotions through speech signals. It finds extensive use in human-computer interaction, virtual assistants, mental health tracking, and automating customer service. This project introduced a deep learning method for SER utilizing Convolutional Neural Networks (CNNs). The system extracts mel-frequency cepstral coefficients (MFCCs) and spectrograms from raw audio inputs and transforms speech signals into two-dimensional images. These images were then processed by a CNN framework that discerns spatial hierarchies and temporal patterns in the speech data to categorize emotions such as happiness, sadness, anger, fear, and neutrality. CNNs improve the capacity of the model to identify local dependencies and invariant features, enhancing its effectiveness in detecting subtle emotional signals. The model was trained and tested on standard datasets such as RAVDESS and CREMA-D to achieve competitive accuracy. This study highlights the potential of CNN-based models for creating robust real-time systems for speech emotion recognition.