AUDIO FEATURE EXTRACTION FOR PREDICTING VIRAL TIKTOK CONTENT USING THE CONVOLUTIONAL NEURAL NETWORK ALGORITHM

Abstract

The rapid growth of TikTok has increased interest in identifying factors that influence content virality, particularly audio elements that play a central role in trend formation and user engagement. This study aims to develop a model for predicting TikTok content virality based on audio characteristics using a Convolutional Neural Network (CNN). The proposed approach focuses exclusively on audio information to evaluate its independent contribution to virality prediction without incorporating visual, textual, or engagement-based features. The research employed a quantitative experimental design using audio extracted from publicly available TikTok videos categorized into viral and non-viral classes. Audio signals were preprocessed through normalization and duration standardization before being transformed into Mel-spectrogram representations. These spectrogram images were then used as input to a CNN model for automatic feature extraction and classification. Model performance was evaluated using precision, recall, and F1-score metrics. The experimental results demonstrate that the proposed CNN model effectively distinguished viral and non-viral TikTok content. Evaluation on the testing dataset produced a precision of 0.833, recall of 1.000, and F1-score of 0.909. In addition, Mel-spectrogram visualizations revealed distinct frequency-energy patterns between viral and non-viral audio samples, indicating that acoustic characteristics contain meaningful information associated with content virality. In conclusion, audio features can serve as reliable predictors of TikTok content virality, and the CNN-based framework successfully extracted discriminative acoustic patterns from Mel-spectrogram representations. This study contributes to the fields of audio analytics and social media intelligence by providing an audio-centered approach for early-stage virality prediction prior to content publication.