FH-RestoreASR: Frequency-Hopping Robust Air Traffic Control Speech Restoration and Recognition

초록

We present FH-RestoreASR, a dropout-aware speech restoration and recognition framework designed for communication systems experiencing short-duration audio dropouts. Such dropouts are common in frequency-hopping systems, widely used for jamming resilience, but they also introduce brief interruptions. These interruptions pose significant challenges for Automatic Speech Recognition (ASR) systems, which typically assume temporally continuous input. FH-RestoreASR introduces a dropout-specific masking mechanism that enables targeted restoration by simulating realistic dropout patterns during training. Built on GAN based architecture, the system restores degraded speech and integrates with ASR models to improve transcription robustness. Experiments on simulated dropout scenarios demonstrate substantial improvements in Word Error Rate (WER) and perceptual quality compared to existing methods. These results highlight the effectiveness of FH-RestoreASR in achieving robust speech recognition under challenging and intermittently disrupted audio conditions.