Method for extracting fatigue features of air-ground communication speech based on cross-validation and recursive features elimination

Purpose Given that air traffic controllers continuously issue air traffic control clearances, this paper aims to introduce a method for extracting speech fatigue features to achieve noncontact fatigue detection, thereby monitoring the controllers’ fatigue status. Design/methodology/approach This paper initially uses five typical methods to extract a variety of features from each instance of air-ground communication speech to achieve a multifaceted description of speech. Subsequently, recursive feature elimination (RFE) and cross-validation are used to evaluate the effectiveness of features, thereby retaining valid fatigue indicators and discarding ineffective ones. The valid fatigue features are then combined to serve as voice fatigue characteristics, and their effectiveness is ultimately validated through experimentation. Findings Compared to methods such as zero crossing rate, Chroma short-time Fourier transform, Mel-frequency cepstral coefficients and root mean square value, the Mel-Spectrogram method yields voice features that more accurately reflect the speaker’s fatigue state. Practical implications This study proposes a method for identifying the fatigue state of air traffic controllers based on air-ground communication speech data. The method is effective in recognizing the fatigue state of controllers and addresses the issue of previous contact-based fatigue detection methods that hindered the work of controllers, thus contributing positively to the safety of civil aviation. Originality/value Despite extensive research, there is currently no universally recognized voice fatigue feature. Addressing this issue, this paper uses RFE and cross-validation to identify a set of features that reflect the speaker’s fatigue state. Experimental results indicate that the number of selected features has been reduced, and their performance surpasses that of other features across various classifiers. This research can serve as a reference for subsequent voice fatigue feature extraction.