FastWave: Optimized Diffusion Model for Audio Super-Resolution

Nikita Kuznetsov; Maksim Kaledin

cs.SDMar 4, 2026

FastWave: Optimized Diffusion Model for Audio Super-Resolution

Authors: Nikita Kuznetsov, Maksim Kaledin

Organizations: HSE University, Russia, St Petersburg, 190121 16 Soyuza Pechatnikov Street · HSE University, Russia, Moscow, 109028 11 Pokrovsky Bulvar

Abstract

Audio Super-Resolution is a set of techniques aimed at high-quality estimation of the given signal as if it would be sampled with higher sample rate. Among suggested methods there are diffusion and flow models (which are considered slower), generative adversarial networks (which are considered faster), however both approaches are currently presented by high-parametric networks, requiring high computational costs both for training and inference. We propose a solution to both these problems by re-considering the recent advances in the training of diffusion models and applying them to super-resolution from any to 48 kHz sample rate. Our approach shows better results than NU-Wave 2 and is comparable to state-of-the-art models. Our model called FastWave has around 50 GFLOPs of computational complexity and 1.3 M parameters and can be trained with less resources and significantly faster than the majority of recently proposed diffusion- and flow-based solutions. The code has been made publicly available.

Figures & tables

Fig. 1: Architecture of FastWave with proposed architectural improvements.

8 kHz
Metric	Baseline 8 NFE	EDM 4 NFE	EDM 8 NFE	FastWave 4 NFE	FastWave 8 NFE
SNR $\uparrow$	$17.47\pm 4.33$	$16.72\pm 4.73$	$16.02\pm 4.60$	$\mathbf{18.49\pm 4.62}$	$18.10\pm 4.31$
LSD $\downarrow$	$1.31\pm 0.12$	$1.25\pm 0.13$	$\mathbf{1.21\pm 0.10}$	$1.22\pm 0.12$	$1.26\pm 0.10$
LSD-LF $\downarrow$	$0.41\pm 0.09$	$0.39\pm 0.09$	$\mathbf{0.32\pm 0.06}$	$0.40\pm 0.07$	$0.36\pm 0.06$
LSD-HF $\downarrow$	$1.42\pm 0.13$	$1.35\pm 0.14$	$1.32\pm 0.11$	$\mathbf{1.31\pm 0.13}$	$1.37\pm 0.11$
12 kHz

TABLE I: Intermediate trained model comparison, the FLOPs are given for one function evaluation.

8 kHz
Metric	FastWave 4 NFE	FastWave 8 NFE	NU-Wave 2 8 NFE	FlowHigh	AudioSR
SNR $\uparrow$	$\mathbf{18.75\pm 4.84}$	$18.53\pm 4.73$	$18.43\pm 4.92$	$18.04\pm 4.74$	$13.75\pm 3.83$
LSD $\downarrow$	$1.18\pm 0.12$	$1.19\pm 0.11$	$1.15\pm 0.10$	$\mathbf{0.96\pm 0.08}$	$1.55\pm 0.15$
LSD-LF $\downarrow$	$0.36\pm 0.08$	$0.28\pm 0.05$	$0.22\pm 0.07$	$\mathbf{0.24\pm 0.02}$	$0.44\pm 0.07$
LSD-HF $\downarrow$	$1.27\pm 0.13$	$1.29\pm 0.12$	$1.25\pm 0.11$	$\mathbf{1.05\pm 0.09}$	$1.69\pm 0.17$
12 kHz

TABLE II: Pretrained / large-capacity model comparison, the FLOPs are given for one function evaluation.

Audio Diffusion Models

Audio diffusion models generate, edit, and restore sound by denoising audio or learned representations under text, audio, or scene conditions.

See all 93 papers

LatentFlowSR: High-Fidelity Audio Super-Resolution via Noise-Robust Latent Flow Matching
Apr 10, 2026Fei Liu, Yang Ai, Hui-Peng Du +2
National Engineering Research Center of Speech and Language Information Processing, University of Science and Technology of China, Hefei, 230027, China
A Survey of Advancing Audio Super-Resolution and Bandwidth Extension from Discriminative to Generative Models
May 15, 2026Ningyuan Yang, Yize Li, Diego A. Cuji +4Audio Diffusion Models Conditional Generative Modeling
Department of Electrical and Computer Engineering Stony Brook University · Department of Electrical and Computer Engineering Northeastern University · Department of Electrical and Computer Engineering University of Illinois Chicago & Discovery Partners Institute
Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion
Mar 7, 2026Maryam Heidari, Nantheera Anantrasirichai, Alin AchimImage Super-Resolution Diffusion Models for Image Restoration
Visual Information Laboratory, University of Bristol, Bristol, UK

Apr 10, 2026J/K move · Enter open · S save

LatentFlowSR: High-Fidelity Audio Super-Resolution via Noise-Robust Latent Flow Matching

Fei Liu, Yang Ai, Hui-Peng Du, Yu-Fei Shi+1

National Engineering Research Center of Speech and Language Information Processing, University of Science and Technology of China, Hefei, 230027, China

Open paper AskPDF arXiv

Audio super-resolution aims to recover missing high-frequency details from bandwidth-limited low-resolution audio, thereby improving the naturalness and perceptual quality of the reconstructed signal. However, most existing methods directly operate in the waveform or time-frequency domain, which not only involves high-dimensional generation spaces but is also largely limited to speech tasks, leaving substantial room for improvement on more complex audio types such as sound effects and music. To mitigate these limitations, we introduce LatentFlowSR, a new audio super-resolution approach that leverages conditional flow matching (CFM) within a latent representation space. Specifically, we first train a noise-robust autoencoder, which encodes low-resolution audio into a continuous latent space. Conditioned on the low-resolution latent representation, a CFM mechanism progressively generates the corresponding high-resolution latent representation from a Gaussian prior with a one-step ordinary differential equation (ODE) solver. The resulting high-resolution latent representation is then decoded by the pretrained autoencoder to reconstruct the high-resolution audio. Experimental results demonstrate that LatentFlowSR achieves competitive or superior performance compared with baseline methods across various audio types and super-resolution settings. These results indicate that the proposed method possesses strong high-frequency reconstruction capability and robust generalization performance, providing compelling evidence for the effectiveness of latent-space modeling in audio super-resolution. All relevant code will be made publicly available upon completion of the paper review process.

Abstract

Audio Diffusion Models

Explore similar work

LatentFlowSR: High-Fidelity Audio Super-Resolution via Noise-Robust Latent Flow Matching

A Survey of Advancing Audio Super-Resolution and Bandwidth Extension from Discriminative to Generative Models

Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion