Automatic Processing Pipeline for Collecting and Annotating Air-Traffic Voice Communication Data

This document describes our pipeline for automatic processing of ATCO pilot audio communication we developed as part the ATCO2 project. So far, collected two thousand hours recordings that either preprocessed transcribers or used semi-supervised training. Both methods using data can further improve by retraining models. The proposed is a cascade many standalone components: (a) segmentation, (b) volume control, (c) signal-to-noise ratio filtering, (d) diarization, (e) ‘speech-to-text’ (ASR) module, (f) English language detection, (g) call-sign code recognition, (h) ATCO—pilot classification and (i) highlighting commands values. key component speech-to-text transcription system has to be trained with real-world ATC data; otherwise, performance poor. In order performance, apply both training contextual adaptation uses list plausible callsigns from surveillance auxiliary information. Downstream NLP/NLU tasks are important an application point view. These need accurate models operating on top real output; thus, there more too. Creating main aspiration At end project, will packaged distributed ELDA.