Following an aircraft accident or incident, investigation entities such as BEA (Bureau of Investigation and Analysis for Civil Aviation Safety) and BEA-é (Bureau of State aviation accident investigations) will systematically analyze flight recorders, commonly called black boxes.

Black boxes are designed to protect flight parameters and cockpit audio recordings in the event of an air accident. The analysis of audio data from cockpit voice recorders (CVR) is absolutely essential in the process of understanding why an accident occured. The duration of CVR recordings has recently moved from 2 hours to 25 hours of audio. The transcription work is particularly time-consuming and greatly affect the workload of investigators trying to establish the circumstances of the accident in order to prevent a similar accident. In this joint effort to evolve their methods and rely on automatic audio data mining algorithms, the BEA and RESEDA have already started to integrate tools, notably automatic speech transcription, with the aim of quickly indexing the content of the recordings and more effectively identify the phases of interest for analysis.

However, the effectiveness of these automatic methods is limited by the presence of a significant quantity of overlapping speech in the CVR recordings. Overlapping speech has two immediate negative impacts. Firstly, it degrades the intelligibility of pilots’ vocal exchanges, and secondly it limits the efficiency of automatic speech transcription tools. Indeed, without the ability to identify the presence of overlapping speech, it becomes difficult to evaluate a posteriori the quality of transcription hypotheses.

In this project, we propose to address the problem of segmentation and separation of superimposed speech with the aim of improving speech intelligibility for human analysis and automatic speech transcription. In particular, the consortium will aim to propose methods for segmentation, automatic intelligibility analysis, and informed source separation. In particular, it will study the contribution that segmentation and automatic evaluation of speech intelligibility can provide to source separation. Furthermore, as part of a frugal, explainable, and reproducible artificial intelligence approach, the methodological framework chosen for source separation will be of the matrix factorization type and will place humans at the heart of the data analysis loop, by initially offering the outputs of simple blind approaches that the AI will inform by providing knowledge if the audio analyst does not judge these first results to be conclusive. At the end of the project, as part of a FAIR procedure, the consortium will aim to make available to the scientific community some data which will be non-sensitive but still representative of the challenges inherent to the CVR.