کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Dual-Branch Cross-Parallel Transformer Model for Single-Channel Speech Enhancement

نویسندگان :

Mohammad Hakimkhah¹ Rahil Mahdian Toroghi² Hassan Zareian³

1- Iran Broadcasting University 2- Iran Broadcasting University 3- Iran Broadcasting University

کلمات کلیدی :

Speech enhancement،Single microphone،Crossparallel Transformer،Dual branch

چکیده :

In this paper, a dual-branch parallel structure is proposed for single-channel speech enhancement, consisting of a Magnitude Mask Branch (MMB) and a Complex Mapping Branch (CMB) utilizing a Cross-Parallel Transformer (CPT) in the time-frequency domain. The CPT effectively captures longterm dependencies along time and frequency axes, extracting time-frequency-related features by integrating their information. The MMB estimates the spectral magnitude, while the CMB compensates for lost spectral details and implicitly extracts phase information. The approach is evaluated on the public VoiceBank+DEMAND dataset. The proposed Dual-Branch Cross- Parallel Transformer Neural Network (DB-CPTNN) achieves superior results compared to SOTA models. Specifically, the model attains PESQ, STOI, SSNR, CSIG, CBAK, and COVL scores of 3.37, 95.9%, 10.58, 4.71, 3.89, and 4.15, respectively, outperforming state-of-the-art (SOTA) benchmarks.