Cockpit background sound separation method based on convolutional spatio-temporal domain audio separation

This paper investigates a cockpit background sound separation method based on a fully convolutional spatio-temporal domain audio separation model. Cockpit background sounds, recorded in the cockpit voice recorder, include warnings, operation noises, and switch sounds. These sounds are critical for evaluating aircraft status and pilot operation. However, due to the diverse types and significant spectral overlap of cockpit background sounds, traditional methods have difficulty achieving efficient separation. This study improves the fully convolutional spatio-temporal domain audio separation network by employing a bidirectional convolutional strategy in the time and channel domains to achieve accurate separation of complex cockpit background sounds. Experiments are conducted using cockpit audio data from Boeing 737- NG aircraft, designing a specialized dataset for validation. Results demonstrate that the proposed method significantly improves separation performance, with SI-SNR and SDR metrics improved by approximately 14.89 dB and 15.13 dB, respectively, compared to the traditional FastICA algorithm, highlighting notable technical advantages.