Speech Emotion Recognition Using Machine Learning and Deep Learning

In today's world, understanding and recognizing human emotions are crucial for how we interact with computers. Exactly finding emotion from speech is a very challenging job. We often use various cues like facial expressions, voice tone, and body movements to identify emotions. This study focuses on Speech Emotion Recognition, which means figuring out emotions from the way people speak. In this model, we have used two datasets the Ryerson Audio-visual Database of Emotional Speech and Song and the Toronto Emotional Speech Set both individually and in a combined dataset. The total number of datasets contains 4048 audio files and nine key emotions happy, calm, angry, sad, neutral, fearful, disgust, surprised, and pleasant surprise. This study employed three feature extraction techniques Mel-frequency cepstral coefficients, Chroma Short-Time Fourier Transform, and Mel spectrograms to generate 180 features from audio files. These features were used to train various classifiers on both the combined and individual datasets. It was found that the Convolutional Neural Network and Multi-Layer Perceptron classifiers outperformed the others on the combined dataset with an accuracy of 83.2 and 84.54 respectively. We believe that this study contributes significantly to human-computer interaction and other applications by enabling more precise emotion recognition.