Cross time–frequency transformer for annoyance evaluation of noise in aircraft cabin

Abstract

Developing a high-precision model for aircraft cabin sound quality prediction is crucial for advancing acoustic design, as it enables reliable prediction-based annoyance prediction rather than relying solely on traditional physical acoustic metrics. However, accurate prediction across various flight phases remains challenging due to noise complexity. Traditional methods (e.g., SVM, XGBoost) and deep learning models (e.g., CNN, LSTM) often fail to fully capture nonlinear relationships or generalize robustly. To address these limitations, this study proposes the Cross Time-Frequency Transformer (CTF-Former), which integrates time-domain and frequency-domain features through a four-stream architecture. The model employs self-attention and cross-attention mechanisms to enhance time–frequency interactions. Additionally, the Self-attention-based Feature Fusion (SaFF) module is introduced to optimize multi-stream feature integration. The CTF-Former demonstrates competitive performance, achieving a normalized MAE of 0.052 ± 0.004 (49.5 % lower than SVR, 25.7 % lower than XGBoost) with reduced variance compared to LSTM (standard deviation decreased by 80 %). It also attains RMSE of 0.068 ± 0.006 and R² of 0.89 ± 0.03 in 5-fold cross-validation, outperforming all compared baselines in average accuracy and stability. The model demonstrates applicability in annoyance prediction and has the potential to support practical applications.