Language Recognition for SSB modulated HF Radio Signals of Short Duration

Language recognition is an important processing step in many speech-related monitoring applications. However, in situations where the quality of the signals is severely degraded, as it is the case for audio signals obtained from radio frequency (RF) transmissions, reliably recognizing the spoken language may be challenging. This is especially true for applications where the speech segments are mostly of short duration. In this work, we focus on language recognition for the demanding case of RF transmissions in the high frequency (HF) band, which are transmitted using single-sideband modulation with suppressed carrier. We apply publicly available state-of-the-art approaches to this task and compare their performances depending on the input speech segment length for a dataset of HF signals recorded in the wild (i.e., natural speech recorded in realistic transmission conditions). Furthermore, a domain adaptation technique is proposed for this HF radio signal scenario and evaluated experimentally.