Evaluation of Mamba Architecture Design Choices for Real-Time Speech Enhancement

Thanks to advances in structured state-space models, which effectively model long sequences with linear computational complexity, Mamba has become a promising alternative to Transformer-based architectures for speech enhancement. However, despite the recent proposal of numerous Mamba-based speech enhancement models, the impact of key architectural design choices on the trade-off between enhancement quality and computational efficiency remains unclear. This paper investigates the computational efficiency of Mamba-based architectures for speech enhancement, focusing on scenarios with limited computational resources. A systematic experimental study was conducted to evaluate the influence of three key architectural factors: the number of Mamba blocks, the embedding dimension, and unidirectional versus bidirectional processing. The models are compared using objective speech enhancement metrics and computational efficiency measures, including model size and inference time. The results provide a comprehensive analysis of the trade-offs between model complexity and enhancement performance.