Speech Emotion Recognition using Mel Spectrogram and Convolutional Neural Networks (CNN)

Humans rely on emotions to express thoughts and engage in everyday activities such as interaction, learning, and decision-making. Speech Emotion Recognition (SER) systems process audio inputs to detect emotions, with applications in sectors like customer service, healthcare, education, and market research. This paper presents a method for SER using Mel spectrograms and Convolutional Neural Networks (CNNs). Audio samples are converted into Mel spectrogram images representing the frequency spectrum on the Mel scale and these images are then used to train a CNN model. The system predicts emotions such as anger, fear, disgust, neutral, happy, surprise, and sad. We evaluate the model using RAVDESS, SAVEE, TESS, and the combined dataset, attaining accuracies of 70% for RAVDESS, 60% for SAVEE, 99.89% for TESS, and 87% for the combined dataset. This approach demonstrates the effectiveness of combining Mel spectrograms with CNNs for robust emotion recognition from speech.