Speech Emotion Recognition in Male Speech Using 58-Dimensional MFCC Features and Random Forest Classification

Abstract

Speech emotion recognition (SER) has developed into a significant research topic in affective computing and human–computer interaction because emotional cues embedded in speech signals can enhance communication between humans and intelligent systems. However, accurately identifying emotional states from speech remains challenging due to dissimilarities in acoustic patterns, speaker features, and recording situations. This study investigates the effectiveness of Mel-Frequency Cepstral Coefficient (MFCC) acoustic features for emotion recognition in male speech using a Random Forest classification model. The dataset used in this research consists of 35,910 male speech samples, each represented by a 58-dimensional MFCC feature vector extracted from emotional speech recordings. The speech samples are categorized into eight emotional classes: angry, fear, calm, disgust, neutral, happy, sad, and surprise. To develop and evaluate the model’s performance, the MFCC data were divided into 80% for training and 20% for testing. The Random Forest model was trained to learn emotional patterns embedded in MFCC features. The experimental findings reveal that the proposed approach achieved an overall classification accuracy of 84.33% with a macro-average F1-score of 0.856, indicating relatively stable performance across multiple emotional categories. Feature importance analysis further reveals that lower-order MFCC coefficients play a dominant role in emotion classification. These findings demonstrate that MFCC features combined with Random Forest classification provide an effective baseline approach for SER and offer valuable insights for future research involving more advanced machine learning models.