کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

A Mamba-Based Symmetric co-Attention Network for Single-Channel Speech Enhancement

نویسندگان :

Mahdi Mazloomi¹ Rahil Mahdian Toroghi² Hassan Zareian³

1- دانشگاه صدا و سیما 2- دانشگاه صدا و سیما 3- دانشگاه صدا و سیما

کلمات کلیدی :

Single channel speech enhancement،GAN،Mamba،feature fusion

چکیده :

In recent years, transformer-based models have advanced speech enhancement through their ability to capture long-range dependencies; however, their quadratic computational cost remains prohibitive for real-time deployment. Conversely, convolution-based methods often struggle to effectively fuse time- and frequency-domain information. This work introduces MambaScA-GAN, a framework that combines the perceptual optimization capability of GANs with the linear-time efficiency of state-space Mamba model. At its core lies the proposed MambaScA module, which leverages a Symmetric co-Attention mechanism to enable dynamic, bidirectional interaction between temporal and spectral features, overcoming the asymmetric fusion limitations of prior architectures. Extensive evaluations on the VoiceBank+DEMAND dataset demonstrate that the proposed model achieves a PESQ of 3.45 and a COVL of 4.21, outper- forming advanced systems such as SE-LMA-Transformer and CMGAN, while requiring less parameters—striking an effective balance between perceptual quality and computational efficiency. ترجمه تخصصی فارسی کن