Robust Speech Micro-Emotion Detection using MFCC Features and CNN in Noisy Environments

Micro-emotions expressed through speech are subtle emotional cues that can be identified from a person's vocal communication. Detecting these emotions is a challenging task because they occur within very short time intervals and often exhibit minimal acoustic variations. This study presents an approach for automatic micro-emotion recognition from speech signals. The proposed framework begins with the application of a Voice Activity Detection (VAD) algorithm to identify and eliminate non-speech portions from the audio recordings. The resulting speech signal is then segmented into smaller units for further analysis. Acoustic features, including Root Mean Square (RMS) Energy, Pitch, Mel-Frequency Cepstral Coefficient (MFCC) Mean, MFCC Variance, and Zero Crossing Rate (ZCR), are extracted from each segment to capture the underlying characteristics of speech.

To improve the effectiveness of emotion classification, the most informative features are selected using the ReliefF feature selection algorithm. These optimized features are subsequently utilized for micro-emotion recognition. The extracted emotional patterns are mapped to corresponding micro-level emotional categories and used to train a Convolutional Neural Network (CNN)-based classification model. The proposed methodology has been evaluated using the CREMA-D (Crowd-sourced Emotional Multimodal Actors Dataset), which contains 7,442 original video clips. Experimental results demonstrate the capability of the system to accurately identify micro-level emotions, even in the presence of background noise, indicating its potential for robust real-world speech emotion recognition applications.