Neural-Network-Based Compressed Sensing for Speech Signals

Some Internet of Things (IoT) applications involving the transmission of speech signals operate in scenarios where sampling is resource-intensive and the transmitter faces severe computational constraints. While Compressed Sensing (CS) provides an efficient framework for compressing and reconstructing sparse signals through sub-Nyquist sampling, traditional CS approaches for audio signals have primarily relied on Matching Pursuit–based algorithms, which are often iterative and computationally demanding. In this paper, we propose two novel neural network–based CS approaches for speech signal processing. Both methods combine spectrogram representations with learning-based reconstruction: the first relies on Autoencoders (AE), while the second leverages recent advances in Generative Adversarial Networks (GANs) combined with meta-learning, known as Compressed Sensing with Meta Learning (CSML). We present experimental results on multiple speech datasets demonstrating that both approaches achieve high reconstruction quality and strong robustness to noise. We further evaluate different network architectures, encoder configurations, and datasets, showing that the proposed methods consistently outperform other CS algorithms commonly considered for IoT applications. The proposed CS-based encoder reflects a hardware-aware engineering trade-off well suited for scalable deployment in resource-constrained IoT environments, while remaining model-free, lightweight, and robust.