Date Approved

6-30-2025

Embargo Period

6-30-2025

Document Type

Thesis

Degree Name

M.S. Electrical and Computer Engineering

Department

Electrical and Computer Engineering

College

Henry M. Rowan College of Engineering

Advisor

Nidhal Bouaynaya, Ph.D.

Committee Member 1

Ying Tang, Ph.D.

Committee Member 2

Tom Tessitore

Abstract

Accurate speech-to-text transcription is critical for Air Traffic Control (ATC) to en- sure safety and operational efficiency. However, existing Automatic Speech Recognition (ASR) systems struggle with domain-specific phraseology, high noise levels, and real-time constraints. This work explores domain adaptation strategies of OpenAI’s Whisper ASR model, applying Transfer Learning and Low-Rank Adaptation (LoRA) on ATC benchmark and a Federal Aviation Administration (FAA) corpus. Experimental results show that fine- tuning significantly reduces Word Error Rate (WER) from 55.2% to 6.8%, yet WER alone fails to capture safety-critical transcription errors. To address WER’s limitations, we introduce the Contextual Error Rate (CER), a novel evaluation metric powered by a Large Language Model (LLM) that prioritizes errors with operational impact rather than surface-level differences. Our experiments showed that while Transfer Learning improved domain-specific accuracy, LoRA preserved better gen- eralization to conversational English. This research demonstrates the importance of combining domain adapted ASR models with context-aware evaluation metrics. By integrating LLM-assisted CER scoring into the evaluation process, we offer a more reliable and context aware method for evaluating tran- scription quality of domain adapted ASR models in aviation, paving the way for scalable, automated speech recognition in air traffic management. The approach proposed is extensible to other context-critical communication systems where transcription fidelity is paramount. These contributions advance the intersection of speech recognition, machine learning, and aviation safety.

Share

COinS