The Observation of Actors’ Vocal Emotion Exercises with Deep Learning and Spectral Analysis

Abstract

This paper presents two distinct methods that demonstrate the increased intensity of a specific emotion when the induced emotion is trained daily for 30 days. For this study, four actors participated in a 30-day exercise trial and were recorded each day using high-level audio equipment. The first method supporting our hypothesis is a deep learning approach. A convolutional neural network pre-trained on Mel-frequency cepstral coefficients analyzed the actors' recordings and delivered the intensity of the detected emotion. The CNN tested 3,561 segments of 0.2-second length, and the results showed a higher level of intensity on the final day of training for each participant. The second method is spectral analysis. The spectrograms generated on the first and final days of the experiment showed that the spectral composition on the final day had a wider range of frequencies than on the first day, further supporting our hypothesis.