Unsupervised Weak Speech Enhancement Using Periodic Mixing Invariant Training

The article delves into the persistent challenge of enhancing weak speech signals in the presence of significant noise, which often leads to a low Signal-to-Noise Ratio (SNR). Traditional methods, including hardware-based amplification and conventional speech enhancement techniques, have limitations in effectively suppressing non-stationary noise and preserving the integrity of speech signals. The proposed framework, TF-Cycle-MixIT, addresses these issues by employing an unsupervised learning approach that leverages the intrinsic harmonic properties and temporal correlations of speech signals. This method autonomously constructs supervised signals without the need for externally labeled training data, making it highly applicable in real-world scenarios where pure speech samples are scarce. The integration of harmonic features with time-domain features enhances the separation of weak speech signals under strong background noise, demonstrating superior robustness in non-ideal acoustic environments. Experimental results, including comparisons with state-of-the-art methods, underscore the effectiveness of TF-Cycle-MixIT in recovering high-fidelity speech signals, even in highly noisy conditions. The article also discusses the implementation details and ablation studies, providing a comprehensive overview of the method's strengths and contributions to the field of speech enhancement.