Automated Extraction of Information from Speech Signals for Speaker Identification and Emotional State Analysis

This study, we discuss the challenging task of acquiring large human speech datasets under constrained conditions, which are typically hand-annotated for speaker recognition and emotional state analysis. We explore a novel approach to learning embedding for speech classification without using labeled audio data. Our hypothesis is that a speech signal can identify a speaker, and by observing the speaker’s facial expressions, we can determine their emotional state. We utilize the VoxCeleb dataset, which contains real-world noise and background chatter, to test our approach. For speaker identification, we apply a CNN-based architecture with the SeLU activation function and Adamax optimizer, achieving an 89.6% accuracy rate. For emotional state analysis, we use a CNN with the Adam optimizer and ReLU activation function, obtaining a 64.97% testing accuracy rate on an external benchmark dataset. Our results suggest that it is possible to infer emotional states from speech signals using facial expression analysis.