Data-Driven Speech Recovery in a Fiber-Optic Polarization-Based Sensing System

Abstract

Optical fibers are inherently sensitive to external acoustic vibrations, which can modulate the local birefringence via the elasto-optic effect, imposing perturbations onto the state of polarization (SOP) of the transmitted light. This creates an unintended sensing channel: for example, speech spoken near the fiber may leak into the SOP trajectory and can potentially be recovered by an eavesdropper. This thesis develops, analyzes, and validates a speech recovery framework that operates directly on SOP obtained from the output of a fiber link. A waveplate fiber channel model is adapted that incorporates the effect of speech on the fiber. Building on this model, a three-stage reproducible speech recovery pipeline is proposed, consisting of preprocessing, demodulation, and enhancement. While confirming their effectiveness, the simulation results show that different demodulation methods give comparable performance, indicating that the primary bottleneck does not lie in the choice of these methods. Building on this insight, hardware experiments are conducted in an optical fiber laboratory using a kilometer-scale single-mode fiber spool as the acoustic sensor. The same pipeline framework used in the simulation study is applied. To further improve the performance, a data-driven speech enhancement method based on a convolutional neural network (CNN) is explored using experimental data, achieving a substantial improvement in perceptual speech quality while preserving intelligibility. Both simulation and experimental results provide consistent support for the fiber channel model, while the experiments further reveal practical performance limitations.