Speech signal quality enhancement based on Mamba and xLSTM architectures

Improving the quality — specifically the intelligibility — of speech signals in the presence of acoustic noise remains a key challenge in the field of digital signal processing. A comparative analysis of neural network-based speech enhancement methods utilizing Mamba and xLSTM architectures is presented. Three configurations are implemented: a baseline model based on MP-SENet, a version incorporating Mamba blocks, and a version utilizing xLSTM blocks. Training and testing are conducted using the VoiceBank+DEMAND dataset, with performance evaluated through objective metrics. Experimental results indicate that the xLSTM-based model demonstrates superior performance across most metrics, including intelligibility (ESTOI = 0.868), spectral accuracy (MCD = 1.98), and perceptual quality (DNSMOS = 3.201), trailing only MP-SENet in terms of the PESQ score. While the Mamba-based model features the fewest parameters (1.3 million), it consistently underperforms compared to the other two approaches in terms of reconstruction quality. The scientific novelty of this work lies in conducting a comprehensive comparative study of Mamba and xLSTM architectures within a unified experimental framework, aimed at enhancing speech signal quality.