Siamese Hybrid RNN Model for Speech Enhancement: A Novel Approach for Noise Reduction in Speech Signals

Efforts to mitigate degradation of speech information and introduction of distortions in speech processing algorithms for smart ha(hearing aid) present a significant challenge. While deep learning-based speech enhancement methods have shown effectiveness, achieving real-time processing remains elusive, particularly for smartphone-centered hearing aid systems. For the purpose of limitation, we suggested a supervised SE approach derived from a Siamese hybrid recurrent neural network (Siamese hybrid RNN) structure. We frame the problem as an approach to enhance speech intelligibility while dealing with limited resources, with a focus on enhancing speech intelligibility in situations with poor SNR scenarios. Described method combines strengths of Siamese architecture and hybrid RNN models to tackle more noise suppression. With the aim of assessing the efficacy of the proposed approach technique, we finalized performance evaluations. Our findings undeniably illustrate the superiority of this approach in enhancing speech intelligibility compared to existing methods. Signal-to-noise ratio has improved by 22%.