MURMUR: Multi-Scale Temporal Learning with Few-Shot Personalization and Explainable AI for Breath-Based Augmentative Communication
Abstract
Voluntary respiratory control is commonly preserved in individuals who have lost speech and limb motor function due to conditions such as amyotrophic lateral sclerosis, locked-in syndrome, and high-level spinal cord injury. This creates a physiologically viable input channel for assistive communication. However, current breath-based Augmentative and Alternative Communication (AAC) systems rely on fixed pressure thresholds, apply population-level classifiers that do not adapt to individual users, and provide no interpretability for clinical validation. This paper presents MURMUR, a system that classifies intentional breath gestures captured by a consumer microphone and adapts to each user from a small calibration set. A Multi-Scale Temporal Convolutional Network (MS-TCN) with four parallel dilated branches at dilation rates d in $\{\text{1, 2, 4, 8} \}$ learns 64-dimensional temporal embeddings from three-channel log-Mel spectrogram features. Evaluated under strict subject-wise splits on the Coswara dataset, the global model achieves macro $\text{F 1}=\text{0. 7 2 6}$ on 382 unseen subjects. Prototypical Networks (ProtoNet) operating on the frozen embedding space personalize the classifier. Cross-subject adaptation at $\mathrm{k}=\text{1 0}$ samples per class reaches macro $\text{F1} =0.730$ without model retraining. An augmentation-based within-subject simulation establishes an upper bound of macro F1 = 0.984 for user-specific calibration. Gradient-weighted Class Activation Mapping (Grad-CAM) generates temporal saliency overlays on the input spectrogram, providing evidence that predictions are influenced by physically interpretable acoustic regions. Ablation shows that multi-scale branching improves macro F1 by 0.013 over a single-scale baseline, while model inference latency remains 2.6 ms. To the authors' knowledge, MURMUR is the first reported breath-based AAC prototype to combine temporal deep learning, few-shot personalization, and explainable AI in a deployed end-to-end pipeline.