Introducing the "Cockpit Party Problem": Blind Source Separation Enhances Aircraft Cockpit Speech Transcription

Puigt M, Bigot B, Devulder H. Introducing the "Cockpit Party Problem": Blind Source Separation Enhances Aircraft Cockpit Speech Transcription. J Audio Eng Soc. 2025;73(1/2):43-53. doi:10.17743/jaes.2022.0189

@article{Puigt2025_22792,
  author = {Puigt, Matthieu and Bigot, Benjamin and Devulder, Hélène},
  title = {{Introducing the "Cockpit Party Problem": Blind Source Separation Enhances Aircraft Cockpit Speech Transcription}},
  journal = {Journal of the Audio Engineering Society},
  volume = {73},
  number = {1/2},
  pages = {43--53},
  year = {2025},
  month = jan,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2022.0189},
  url = {https://doi.org/10.17743/jaes.2022.0189}
}

TY  - JOUR
TI  - Introducing the "Cockpit Party Problem": Blind Source Separation Enhances Aircraft Cockpit Speech Transcription
AU  - Puigt, Matthieu
AU  - Bigot, Benjamin
AU  - Devulder, Hélène
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 73
IS  - 1/2
SP  - 43
EP  - 53
PY  - 2025
DA  - 2025/01/07
DO  - 10.17743/jaes.2022.0189
UR  - https://doi.org/10.17743/jaes.2022.0189
PB  - Audio Engineering Society
LA  - en
AB  - Cockpit voice recorders (CVRs) are one of the two mandatory flight recording devices embarked in commercial aircraft. Its analysis is crucial to understand the context of an air incident or accident. However, in such scenarios, when the audio recordings are usable, CVR may contain strong mixtures of crew member speech signals, radio communications, and cockpit alarms. However, contrary to the “cocktail party problem” that blind source separation (BSS) aims to tackle, modeling CVR mixtures—that are here named the “cockpit party problem”—was never done before. In this paper, the authors thus propose a CVR mixture model and highlight its limitations. While not trivial—even in a two-source scenario—BSS methods can be applied to real CVR recordings. It is found that taking into account several BSS outputs provided by various methods may help audio analysts to transcribe the CVR data. That is, near 90% of unintelligible words can be transcribed from CVR recordings processed by BSS methods.
ER  -