Emotion Detection From Speech for Improved Communication Accessibility

This paper aims to develop a robust emotion recognition system that accurately classifies emotions from speech, with potential applications in accessibility technologies, gauging mental health, and interactions among users and computers. The primary objective is to devise a model that is adept at recognizing the elementary emotions-angry, disgust, fear, happy, neutral, sad, and surprise-using audio features extracted from speech datasets. Through the exploitation of the diversity of datasets available, that include, TESS, RAVDESS, and CREMA-D and SAVEE, we ensure that the model is highly generalizable across the different variations of speakers, emotional intensities, and recording environments. We preprocess the data by unifying emotion labels, standardizing formats, and extracting prosodic features like Mel-Frequency Cepstral Coefficients (MFCCs). After preparing the data, we apply machine learning techniques to capture spatial and temporal features based on the voice signals. The model is trained and evaluated on an 80/20 split, ensuring it is tested on a wide variety of unseen data. The model reaches an accuracy rate of 95%, with balanced precision, recall, and F1-scores across all emotion categories. It exhibits strong performance in categorizing emotional states from speech. Its potential is considerable in practical scenarios, especially in improving accessibility for the deaf and hard-of-hearing, enriching user experiences using AI-driven systems, and even in assistance with diagnostics related to mental health.