속도와 노이즈 증강 데이터를 통한 음성 감정 인식 성능 향상

Abstract

Speech Emotion Recognition Performance Improvement with Speed and Noise Augmentation Data Ahn, JinSung Advisor : Prof. Shin, JuHyun Ph.D. Department of Software Convergence Engineering Graduate School of Industrial Technology and Entrepreneurship, Chosun University With the recent development of AI in the field of voice emotion recognition research is expanding to utilize the characteristics of emotions revealed in human voices through tone of voice or non-verbal elements. However, the voice data sets currently distributed are limited, and there were practical difficulties in collecting the desired data sets directly. Therefore, this study proposed an experiment to improve voice emotion recognition by applying speed and noise enhancement methods and MFCC and Mel-Spectrogram feature vector extraction techniques based on a small amount of data sets. The study used CREMA datasets classified into a total of five emotional classes: 'anger', 'fear', 'happiness', 'sad', and 'neutral'. Data was augmented by adjusting speed and noise expansion, and vectors were extracted using two feature vector extraction techniques, MFCC and Mel-Spectrogram. As a result of the experiment, data augmentation was applied rather than the accuracy of learning only the original data, and the accuracy of learning resulted in a performance improvement of about 3% or more.