# VM-ASR: A Lightweight Dual-Stream U-Net Model for Efficient Audio Super-Resolution

- URL: https://archi-zhang.blog/portfolio/vm-asr-a-lightweight-dual-stream-u-net-model-for-efficient-audio-super-resolution
- Type: Publication
- Year: 2024

A novel audio super-resolution model published in Q1 IEEE TASLP (2025), achieving 35% SOTAimprovement in Log-Spectral Distance with only 3.01M parameters.

**VM-ASR** is a lightweight, dual-stream **U-Net** model designed for efficient **audio super-resolution (ASR)**, or bandwidth extension. It enhances low-resolution audio (e.g., 8 kHz) by reconstructing missing high-frequency components to produce high-fidelity sound (e.g., 48 kHz). The model integrates **Visual State Space (VSS)** blocks derived from **VMamba** to capture both global and local acoustic contexts, while its dual-stream U‑Net architecture separately processes magnitude and phase spectra to improve harmonic accuracy and phase reconstruction.

Extensive evaluation on the **VCTK dataset** shows that VM‑ASR surpasses state‑of‑the‑art approaches in spectral reconstruction quality across multiple upscaling configurations. Despite its strong performance, it maintains an exceptionally compact design—just **3.01 M parameters** and **2.98 GFLOPS**—achieving **~27× real‑time** processing speed for 16 kHz → 48 kHz upsampling on a Tesla V100 GPU. This combination of accuracy, speed, and efficiency highlights VM‑ASR’s potential for **real‑time deployments** in **telecommunications, speech synthesis, and audio restoration**, enabling high‑quality sound enhancement even in **resource‑constrained environments**.

[GitHub](https://github.com/ghnmqdtg/vm-asr) • [Paper](https://ieeexplore.ieee.org/document/10852332) • [Demo](https://ghnmqdtg.github.io/vm-asr-demo/)