Lung-SRAD: Spectral-Aware Regularized Audio DASS with Dual-Axis Patch-Mix Contrastive Learning for Respiratory Sound Classification

2026-06-10 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, AI in Healthcare · Depth: Expert, quick

Summary

Lung-SRAD introduces a novel approach for Respiratory Sound Classification (RSC) that leverages State Space Models (SSMs) as an alternative to traditional CLS-token driven self-attention architectures like the Audio Spectrogram Transformer (AST). Existing AST models often exhibit low-pass filtering, diminishing sensitivity to localized abnormal respiratory patterns. Lung-SRAD addresses this by utilizing SSMs, which demonstrate stronger preservation of mid-to-high spatial-frequency components in intermediate representations. The method incorporates spectral-aware layer regularization via Gaussian convolution and proposes Dual-Axis Patch-Mix contrastive learning specifically for SSM-based audio models. This combined strategy achieved a 64.48% score on the ICBHI benchmark, surpassing the AST baseline by 5%. Code is publicly available.

Key takeaway

For Machine Learning Engineers developing respiratory sound classification systems, Lung-SRAD offers a compelling alternative to AST models. If your current models struggle with localized abnormal patterns due to low-pass filtering, you should investigate integrating State Space Models and spectral-aware regularization. This approach, which achieved 64.48% on ICBHI, suggests a path to significantly improve diagnostic accuracy by better preserving critical mid-to-high frequency audio details. Consider exploring the provided code to adapt these techniques.

Key insights

Lung-SRAD uses State Space Models and spectral-aware regularization to improve respiratory sound classification by preserving high-frequency details.

Principles

• SSMs preserve mid-to-high spatial-frequencies.
• Spectral-aware regularization enhances sensitivity.
• Contrastive learning improves representation robustness.

Method

Lung-SRAD employs Distilled Audio State Space models, applies Gaussian convolution for spectral-aware layer regularization, and integrates Dual-Axis Patch-Mix contrastive learning for robust representation.

In practice

• Use SSMs for high-frequency audio tasks.
• Apply Gaussian convolution for spectral regularization.
• Implement Dual-Axis Patch-Mix contrastive learning.

Topics

• Respiratory Sound Classification
• State Space Models
• Contrastive Learning
• Spectral-Aware Regularization
• Audio Spectrogram Transformer
• ICBHI Benchmark

Code references

• RSC-Toolkit/Lung-SRAD

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Contrastive Training with LLM-generated Near-Misses for Robust Code-Switching Speech Recognition — Contrastive training with LLM-generated near-misses robustly improves code-switching ASR at critical language-alternation points.
• Quality Adaptive Angular Margin Learning for Respiratory Sound Classification — QLung uses quality-adaptive angular margins to improve respiratory sound classification generalization and handle class imbalance.
• Echo: A Joint-Embedding Predictive Architecture for Speaker Diarization and Speech Recognition in a Shared Latent Space — A single 25M-parameter ViT encoder can jointly perform speaker diarization, speech recognition, and source separation in a shared latent space.
• Speech-Guided Multimodal Learning for Vocal Tract Segmentation in Real-Time MRI — Multimodal supervision during training can enhance single-modality inference for challenging medical image segmentation tasks.
• Diffusion Attention Expert Model for Predicting and Semi-automatic Localizing STAS in Lung Cancer Histopathological Images — DAEM accurately detects STAS in lung cancer histopathology, offering strong generalizability and TME-based localization.
• AST: Adaptive, Seamless, and Training-Free Precise Speech Editing — AST enables training-free, precise speech editing by manipulating latent spaces of pre-trained autoregressive TTS models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.