Grad-CAM-Guided Curriculum Learning for Band-Wise Blind Estimation of Room-Acoustic Parameters from Noisy Speech

This work addresses blind estimation of octave-band reverberation time (T60) and clarity (C50) from noisy reverberant speech. We propose a Grad-CAM-guided Curriculum Learning (CL) strategy for a Vision Transformer (ViT)-based multi-head regressor. The method applies octave-centered time-frequency masks before the backbone and progressively relaxes them during training, encouraging band-consistent representations while preserving full-spectrum context. Experiments compare ResNet-50 and ViT-B/16 baselines under synthetic and measured room impulse response (RIR) conditions. The optimized six-head ViT-B/16 model with CL (M3+CL) improves low-frequency estimation and provides a favorable accuracy-efficiency trade-off for short speech blocks. At L = 3.0 s, M3+CL remains competitive with recent attention-based estimators and yields more localized, physically consistent Grad-CAM attention maps.