Authors - Roberts Dargis, Arturs Znotins, Ilze Auzina, Maris Golubovskis, Mikelis Gulbis, Normunds Gruzitis Abstract - Operational radio communication is a challenging application domain for automatic speech recognition (ASR) despite advances in multilingual foundation models. We investigate the applicability of modern Latvian ASR models to operational communication and evaluate whether speech enhancement techniques improve recognition quality under realistic conditions. To support the study, we created a specialized corpus of authentic Latvian operational radio communication. The corpus captures acoustic and linguistic phenomena largely absent from general speech corpora, including narrow-band transmission, radio-channel artifacts, environmental noise, domain-specific terms, and fragmented utterances. Using this corpus, we evaluate state-of-the-art adaptations of the massively multilingual Whisper and MMS models in combination with several audio preprocessing methods. The results reveal a substantial performance gap between the conventional Latvian ASR benchmarks and operational communication data. While some preprocessing methods improve perceived audio quality, they provide limited benefit for downstream recognition and often even degrade ASR performance. Voice activity detection, however, yields the most consistent improvements. The findings indicate that domain mismatch, rather than acoustic degradation alone, is the dominant source of recognition errors and highlight the need for representative domain-specific data when adapting general-purpose ASR models for the operational communication environment.
Open Zoom