Deep Causal Speech Enhancement and Recognition Using Efficient Long-Short Term Memory Recurrent Neural Network

Abstract - In this work, we propose an attention-based beamforming framework for multi-channel speech enhancement that dynamically adapts spatial filtering to complex acoustic environments. Traditional beamforming methods rely on fixed or heuristically derived spatial filters, limiting their robustness in the presence of non-stationary noise and reverberation. Our approach leverages a self-attention mechanism to learn context-aware representations of spatial cues across multiple microphone channels, enabling the model to emphasize target speech while suppressing interfering sources. By integrating the attention mechanism within a neural beamformer architecture, we enable end-to-end optimization of both spatial filtering and spectral enhancement. Experiments conducted on benchmark multi-channel datasets demonstrate significant improvements in signal-to-noise ratio (SNR), perceptual quality (PESQ), and speech intelligibility (STOI), outperforming conventional and deep learning-based baselines. This method offers a promising direction for robust speech enhancement in real-world far-field and noisy scenarios.Multi-channel speech enhancement aims to extract clean speech from noisy and reverberant environments using spatial information captured by multiple microphones. Conventional beamforming methods, such as Minimum Variance Distortionless Response (MVDR) and Generalized Eigenvalue (GEV) beamformers, typically rely on accurate estimation of spatial covariance matrices (SCMs) and steering vectors, which are challenging to compute in dynamic or real-world conditions. Key Words: Multi-channel speech enhancement Attention mechanism Beamforming Neural beamformer Self-attention