Full-Band Audio Super-Resolution Using Generative Adversarial Networks
Silaev, Mikhail (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-05-15
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202505155522
https://urn.fi/URN:NBN:fi:tuni-202505155522
Tiivistelmä
Generative learning, which enables the synthesis of novel and authentic content from existing data, represents a highly promising direction in artificial intelligence (AI). This thesis focuses on a specific challenge within this domain: audio super-resolution using generative adversarial networks (GANs).
The task involves increasing the sampling rate of digital audio signals while extending their bandwidth by generating high-frequency spectral components. A GAN-based model is developed to upsample audio from 4 to 16 kHz and from 16 to 48 kHz. The latter includes bandwidth extension up to 24 kHz, referred to as full-band super-resolution, as it surpasses the upper limit of typical human hearing. Specific strategies to achieve high perceptual quality are proposed to suppress common upsampling artifacts, such as checkerboard patterns.
The generative model developed in this thesis produces high-fidelity full-band audio that closely resembles wideband reference signals. It outperforms several existing methods in subjective listening tests. These results demonstrate the feasibility of full-band audio super-resolution with generative AI, establishing a state-of-the-art approach for generating high-quality 48 kHz audio.
The task involves increasing the sampling rate of digital audio signals while extending their bandwidth by generating high-frequency spectral components. A GAN-based model is developed to upsample audio from 4 to 16 kHz and from 16 to 48 kHz. The latter includes bandwidth extension up to 24 kHz, referred to as full-band super-resolution, as it surpasses the upper limit of typical human hearing. Specific strategies to achieve high perceptual quality are proposed to suppress common upsampling artifacts, such as checkerboard patterns.
The generative model developed in this thesis produces high-fidelity full-band audio that closely resembles wideband reference signals. It outperforms several existing methods in subjective listening tests. These results demonstrate the feasibility of full-band audio super-resolution with generative AI, establishing a state-of-the-art approach for generating high-quality 48 kHz audio.