Prediction of Electric Toothbrush Acoustic Comfort Using a Lightweight Dual‐Branch CNN With Psychoacoustic‐Spectral Cross‐Attention Fusion

ABSTRACT

Acoustic comfort has become a critical factor influencing user experience in electric toothbrushes. However, traditional subjective evaluation methods are time‐consuming, labor‐intensive, and difficult to scale for rapid product iteration. This study proposes a lightweight dual‐branch convolutional neural network with psychoacoustic‐spectral cross‐attention fusion (LDCNN‐CF) to predict perceived acoustic comfort by integrating Mel‐spectrograms and psychoacoustic parameters. A dataset of 600 audio samples (8 brands × 3 modes × 5 scenarios × 5 repetitions) was constructed, with subjective comfort scores obtained from 80 listeners using the rank score comparison (RSC) method. Through 8‐fold leave‐one‐brand‐out cross‐validation, the LDCNN‐CF model achieves an MAE of 0.82 and R ² of 0.84 using only 0.42 million parameters. It consistently outperforms both traditional machine learning models and state‐of‐the‐art deep learning baselines. Ablation studies confirm the complementary contributions of spectral and psychoacoustic features, while correlation and attention analyses identify roughness as the dominant predictor of acoustic discomfort ( ρ  = −0.79). The model also demonstrates strong generalization across unseen brands (MAE ranging 0.78–0.85). The proposed framework offers an accurate, efficient, and interpretable solution for data‐driven sound quality evaluation, with strong potential for on‐device deployment in resource‐constrained consumer product design workflows.