Speech Emotion Recognition Project Classifies Six Emotions from Audio Using RAVDESS

A new open project explores how machine learning can detect emotional cues in speech by analyzing pitch, energy, rhythm, and tone. It is trained on the RAVDESS dataset of acted emotional speech and classifies clips into six labels: neutral, happy, sad, angry, fearful, and disgust. Audio is standardized and converted into MFCC features and log-mel spectrograms, with models including SVM, random forest, XGBoost, and a small CNN compared on validation and test data. Speaker-independent splits are used so the same voice does not appear across training and evaluation, and a FastAPI service exposes predictions through an upload endpoint. The authors stress that outputs reflect patterns in acted data rather than a reliable reading of a speaker's true emotional state.