Title: High quality video creation that integrates audio and visual feature information

Authors: Hua Fan; Dongdong Lin

Addresses: Department of Cinematography, Beijing Film Academy, Beijing, 100088, China ' Department of Cinematography, Beijing Film Academy, Beijing, 100088, China

Abstract: Video generation involves creating virtual video images through computer programs with diverse applications. This study focuses on cross-modal video generation from audio to visual information. It extracts Mel cepstral coefficient features from audio signals and establishes a convolutional neural network with self-attention. A high-quality video creation model is developed based on neural radiation field representation. Results indicate an Inception score around 1.6 to 1.7, showing good convergence of the loss function at 40,000 training frequency. The proposed algorithm enhances peak signal-to-noise ratio by 24.7% compared to standard methods, with training speed improved by up to 1-2 times. The model exhibits resolution flexibility, robustness across scenarios, and improved video viewing experience compared to existing methods, promising advancements in video creation technologies.

Keywords: audio features; visual features; video creation; Mel cepstral coefficient characteristics; convolutional neural network; CNN.

DOI: 10.1504/IJWET.2026.153359

International Journal of Web Engineering and Technology, 2026 Vol.21 No.1, pp.27 - 45

Received: 28 Apr 2024
Accepted: 18 Mar 2025
Published online: 06 May 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article