An array speech enhancement algorithm based on the Hippopotamus Optimization algorithm and Mamba

Abstract Microphone arrays often suffer from robustness degradation and performance loss in complex acoustic environments due to suboptimal geometric design and inadequate enhancement algorithms. This paper proposes a novel two-stage optimal design method—comprising (i) geometric optimization of the microphone array for robustness, and (ii) end-to-end speech enhancement via a deep learning model. First, the Hippopotamus Optimization (HO) algorithm is leveraged to optimize the array geometry, yielding a physically more robust layout (quantified by a high White Noise Gain (WNG) of 7.29 dB—a 5.48 dB improvement over other metaheuristics). Second, an end-to-end enhancement model based on the Mamba state-space network is introduced; chosen for its linear-time complexity, it directly maps multi-channel noisy input to clean speech without requiring error-prone source localization. Extensive experiments show the integrated system’s superiority: the Mamba enhancer achieves an average SI-SNR gain of 5.02 dB over DSB, 13.22 dB over MVDR, 1.90 dB over LSTM, and 1.38 dB over Transformer, with pronounced advantages in low-SNR conditions. This work demonstrates that co-optimizing physical array design and deep learning-based processing significantly advances microphone array performance for real-world applications. As a critical front-end component in acoustic signal processing, the performance of microphone arrays largely depends on the geometric configuration of the array and the effectiveness of speech enhancement algorithms. However, in complex acoustic environments, microphone arrays often suffer from performance degradation, manifested as insufficient robustness and poor enhancement performance. To address this issue, this paper proposes a two-stage optimal design method. Specifically, the Hippopotamus Optimization algorithm (HOA) is utilized to obtain the geometric configuration of the microphone array with high robustness and low coherent noise, and the Mamba algorithm is employed for speech enhancement. Unlike traditional beamformers that rely on accurate source localization, the proposed Mamba-based enhancement network learns the mapping from noisy to clean speech directly, making it robust in dynamic environments where the source position may be unknown. Compared to traditional methods, the proposed algorithm demonstrates superior performance in terms of speech quality and noise suppression, especially in low signal-to-noise ratio environments.The code is available at https://github.com/ztysth/SpeechEnhance_Mamba .