Audio super-resolution system for converting narrowband audio into full-band audio

Background

Bandwidth extension, or audio super resolution, is a signal processing technique used to reconstruct missing high-frequency content in audio recordings. The goal is to convert narrowband signals into full-band audio sampled at 44 kHz or higher. There is a significant need for this technology in the modern media landscape. Today, much audio is captured using consumer electronics like smartphones, Bluetooth headsets, and web conferencing software. These devices inherently compress and degrade audio quality, rendering the recordings unsuitable for professional audio and video production where pristine, high-fidelity sound is required.

Despite this demand, current approaches to bandwidth extension suffer from severe limitations. Traditional methods fail to accurately recreate complex high-frequency spectrums. While deep learning techniques have attempted to bridge this gap, they predominantly target maximum sample rates of only 16 to 22 kHz, falling drastically short of professional standards. Furthermore, existing neural network models frequently generate overly smoothed spectrograms that lack crucial fine acoustic details. These conventional systems also struggle with accurate phase reconstruction and rely on basic loss functions, which inevitably introduce unnatural, audible artifacts into the final signal, preventing true high-fidelity reproduction.

Technology

This audio super-resolution system utilizes a generative adversarial network (GAN) to convert narrowband audio into full-band audio at sample rates of 44 kHz or higher. The solution upsamples input audio and passes it through a feedforward WaveNet generator employing non-causal dilated convolutions to estimate missing high-frequency components. To ensure exceptional quality, the system incorporates multiple discriminators, including a spectrogram discriminator and several waveform discriminators operating at varying downsampled rates. For efficient execution, the technology divides audio into parallelized sub-batches, which are reassembled using phase-preserving stitching to maintain temporal and spectral consistency across single and multichannel signals.

This technology is differentiated by its ability to surpass traditional bandwidth extension methods, which typically cap at 22 kHz and produce smoothed spectrograms with noticeable artifacts. By targeting professional-grade sample rates up to 48 kHz, it allows audio captured on standard consumer devices to be used in high-end production. The solution achieves this through a unique composite loss function—combining adversarial, feature matching, and multi-resolution spectral losses—that balances realism and fidelity. Additionally, the model is trained with noise-augmented data, making it highly robust against real-world recording artifacts and rendering the generated audio perceptually indistinguishable from authentic high-resolution recordings.

Benefits

• Transforms low-quality, narrowband audio from consumer devices into professional-grade, high-fidelity full-band audio (44 kHz or higher).
• Generates highly realistic missing high-frequency content that is perceptually indistinguishable from actual high-resolution recordings.
• Provides robust performance against real-world background noise and recording artifacts.
• Ensures fast and efficient processing through parallelized batch execution while maintaining seamless quality via phase-preserving stitching.
• Supports both single-channel and complex multichannel audio formats, including stereo and surround sound.
• Offers flexible deployment options, allowing for easy integration into digital audio workstations (DAWs), cloud services, plugins, or standalone applications.

Applications

• Audio post-production enhancement
• Archival audio recording restoration
• Video conferencing audio upscaling
• Podcast audio quality enhancement
• Smartphone voice recording upscaling

Additional info

This GAN-based framework synthesizes full-band audio (≥44 kHz) from narrowband inputs. A feedforward WaveNet generator utilizes non-causal dilated convolutions to reconstruct high-frequency data. The architecture integrates multi-scale waveform and 2D-convolutional spectrogram discriminators. Training leverages composite adversarial, feature-matching, and multi-resolution spectral losses. Enhanced by noise augmentation and phase-preserving sub-batch stitching, the system ensures temporal and spectral consistency for high-fidelity, multichannel signal reconstruction.