Two-Stage Cascaded Speech Enhancement by Exploiting Magnitude and Phase Optimization

This article delves into the critical area of speech enhancement, addressing the persistent challenges of improving speech quality and intelligibility in noisy environments. Traditional methods, while effective, often struggle with the mismatch problem, leading to significant performance degradation. The proposed two-stage cascaded speech enhancement method offers a novel solution by separately optimizing the magnitude and phase components of speech signals. The first stage focuses on recovering the magnitude using an estimated mask network, while the second stage optimizes the phase, leveraging the complementary information between magnitude and phase. This approach not only enhances the overall performance of speech enhancement but also provides a clearer learning objective for the network, improving its interpretability. The article also introduces an implicit phase optimization architecture to avoid the interference caused by direct phase estimation, further refining the speech enhancement process. Through extensive simulations and comparisons with baseline methods, the proposed method demonstrates superior performance in terms of perceptual evaluation of speech quality (PESQ), short-time objective intelligibility (STOI), and signal-to-distortion ratio (SDR), making it a significant contribution to the field of speech enhancement.