Institutional Repository of Vilnius University
Triukšmui atsparus AM radijo kalbos atpažinimo modelis su maršrutizavimu sąlygotais LoRA adapteriais /
Abstract
dc:descriptionThis work presents a hybrid system for automatic speech recognition (ASR) in aviation very high frequency (VHF) radio communications. A novel methodology is introduced, combining an ECAPA-TDNN classification model used as an audio routing component with four distinct Whisper-Small LoRA adapters, each conditioned on specific noise pseudo-labels. This architecture allows even a relatively small model to achieve competitive transcription accuracy. Compared to the baseline Whisper-Small, Whisper-Large models, and a non-conditional LoRA adapter, the proposed system achieved word error rate (WER) reductions of 77~\%, 66.65~\%, and 5.76~\% respectively. Compared to existing specialized models, the proposed system trails by only 9.78~\% points while being 6.4 times smaller. Additionally, a methodology for synthesizing VHF amplitude modulation (VHF AM) radio data using open-source tools was developed.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sabaitis, Rokas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308334924&prefLang=en_US