Analyzing the Impact of HF-Specific Signal Degradation on Automatic Speech Recognition

Analog radio still constitutes an important fraction of military communications. In the context of a fully digitized radio frequency (RF) reconnaissance chain, this leads to potentially massive amounts of digitized analog modulated speech data that has to be processed. A crucial step in information extraction from this speech data is automatic speech recognition (ASR), i.e., speech-to-text conversion. Whereas modern deep learning methods have immensely advanced state-of-the-art ASR performance in the last years, applicability to reconnaissance scenarios is not immediate. In this paper we provide a systematic analysis of a state-of-the-art ASR system applied to the demanding scenario of amplitude modulated speech transmissions over the high frequency (HF) radio channel. We evaluate ASRrobustness in terms of word and character error rate for various types of channel conditions, distortions and adversary effects such as carrier frequency misestimation. We thus derive valuable insights on both how to integrate modern ASR solutions in an automatic reconnaissance chain and which limitations of such systems are yet to be overcome. This paper was originally presented at the NATO Science and Technology Organization Symposium (ICMCIS) organized by the Information Systems Technology (IST) Panel, IST-20S-RSY- the ICMCIS, held in Koblenz, Germany, 23-24 April 2024.