A Deep Learning Framework for Real-Time Speech Emotion Recognition

Speech Emotion Recognition (SER) is the task of estimating a speaker’s emotion from their voice. This study compares three ways of representing speech using Mel-Frequency Cepstral Coefficients (MFCCs): 13, 26, and 39 coefficients. The features are supplied to a Stacked Gated Recurrent Unit (GRU) model, which learns patterns that change over the course of an utterance. The experiments use 4,240 audio samples from the RAVDESS and Toronto Emotional Speech Set (TESS) datasets, grouped into seven emotion classes. Each MFCC sequence is padded or shortened to 200 frames, and feature scaling is fitted using the training data only. The three models use the same architecture and training settings so their results can be compared on a consistent basis. Evaluation in the project includes classification metrics, one-vs-rest ROC-AUC, probability calibration curves, and t-SNE plots. The Phase-II report identifies MFCC-39 as the most promising of the three variants and MFCC-13 as a lighter baseline. Since the report’s summary table labels its accuracy and F1 values as expected performance, those estimates should not be presented as confirmed test results unless supported by the final experimental outputs.