Fine-Tuning Whisper for American English Air Traffic Control Speech Recognition: A Data-Efficient Pipeline

Automatic speech recognition (ASR) for air traffic control (ATC) presents a persistent domain adaptation challenge: VHF radio channel degradation, specialized vocabulary, and a near-total absence of American English training corpora in the published literature. The state-of-the-art open ATC ASR model—WhisperATC by van Doorn et al. [20], fine-tuned on European ATCO2 and ATCOSIM corpora—achieves 3.88% word error rate (WER) on ATCOSIM (speaker-split evaluation) but degrades to 30.3% on American ATC transmissions, exposing a systematic accent and phraseology mismatch. We present a data-efficient fine-tuning pipeline that adapts Whisper Large v3 [17] to American English ATC using only 55 manually transcribed clips recorded from three major US airports (KIAH, KJFK, KSFO) via LiveATC.net. Domain-matched audio preprocessing—a 300–3400 Hz Butterworth bandpass filter with EBUR128 loudness normalization—combined with five-fold stochastic data augmentation addresses the limited corpus size. Full fine-tuning with conservative hyperparameters achieves 13.7% WER, a 54.8% relative reduction from the European-trained baseline, using 370× fewer training clips than the most comparable prior study. A secondary contribution is the characterization of a structural incompatibility between the HuggingFace PEFT LoRA implementation and Whisper’s log-mel spectrogram encoder that prevents parameter-efficient fine-tuning without modification of library internals. All code, the fine-tuned model, and training notebooks are publicly available.