Improving speech recognition models with small samples for air traffic control systems

Abstract In the domain of air traffic control (ATC) systems, efforts to train a practical automatic speech recognition (ASR) model always faces problem small training samples since collection and annotation are expert- domain-dependent task. this work, novel approach based on pretraining transfer learning is proposed address issue, an improved end-to-end deep developed specific challenges ASR in ATC domain. An unsupervised strategy first learn representations from unlabeled for certain dataset. Specifically, masking applied improve diversity sample without losing their general patterns. Subsequently, fine-tune pretrained or other optimized baseline models finally achieves supervised By virtue common terminology used domain, task can be regarded as sub-domain adaption task, which transferred using joint corpus consisting new transcribed target This construction enriches size samples, important addressing issue corpus. addition, speed perturbation augment further quality Three real datasets validate strategies. The experimental results demonstrate that performance significantly all three datasets, with absolute character error rate only one-third achieved through training. applicability strategies approaches also validated.