DBRNN: Robust Speech Bandwidth Expansion Via High Fidelity Dilated Bidirectional Recurrent Neural Network

Speech Bandwidth Extension (SBE) is critical for enhancing speech quality and intelligibility in bandwidth-constrained telecommunications, where limited frequency range often leads to degraded audio clarity. However, state-of-the-art algorithms face challenges such as inadequate high-frequency signal reconstruction, limited robustness in noisy conditions, and inefficiencies in capturing complex temporal dependencies. To address these, this study introduces a Dilated Bidirectional Recurrent Neural Network (DBRNN)-based SBE framework, organized into Encoder, Bottleneck, and Decoder stages. The bottleneck integrates Intra-chunk and Inter-chunk processing techniques, effectively modeling long and short-term sequences to optimize high-frequency signal reconstruction. Evaluations on IEEE datasets reveal that the proposed approach outperforms baseline methods, achieving significant improvements in metrics like STOI, PESQ, LSD, and SOPM, while enhancing speech naturalness and intelligibility. By leveraging advanced neural network architectures, this work addresses the limitations of traditional algorithms, offering a robust and efficient solution for enriching communication experiences across diverse platforms and environments.