A lightweight end-to-end speech restoration framework for multi-degradation environments

In this work, we propose a lightweight and modular speech restoration framework that addresses three common types of audio degradation: bandwidth limitation, additive noise, and reverberation. The proposed framework employs a common encoder–BLSTM–decoder architectural design that is instantiated separately for each restoration task. Each task-specific model is trained independently for its corresponding degradation condition while retaining the same network structure and spectrogram-domain processing pipeline. Operating in the log-magnitude spectrogram domain enables high-resolution time–frequency analysis, while the Bidirectional Long Short-Term Memory (BLSTM) bottleneck captures long-range temporal dependencies within each task-specific model. All models are optimized using a multi-resolution STFT loss comprising spectral convergence and log-magnitude distance components, thereby encouraging spectral fidelity during restoration. Performance is evaluated using PESQ, STOI, and SI-SDR on degraded–clean speech pairs generated under controlled bandwidth limitation, additive-noise, and room-reverberation conditions. Despite containing only 4.8M parameters, substantially fewer than the compared architectures such as DCHT (13.0M), MP-SENet (15.6M), and CMGAN (17.2M), the proposed framework achieves competitive restoration performance, with PESQ scores in the range of 2.95–3.05 and a STOI of approximately 0.86. The compact architecture indicates its potential suitability for resource-constrained speech restoration applications.