Adaptive Test-Time Scaling for Zero-Shot Respiratory Audio Classification

Tsai-Ning Wang; Herman Teun den Dekker; Lin-Lin Chen; Neil Zeghidour; Aaqib Saeed

cs.SDApr 14, 2026

Adaptive Test-Time Scaling for Zero-Shot Respiratory Audio Classification

Authors: Tsai-Ning Wang, Herman Teun den Dekker, Lin-Lin Chen, Neil Zeghidour, Aaqib Saeed

Organizations: Eindhoven Artificial Intelligence Systems Institute, The Netherlands

Abstract

Automated respiratory audio analysis promises scalable, non-invasive disease screening, yet progress is limited by scarce labeled data and costly expert annotation. Zero-shot inference eliminates task-specific supervision, but existing methods apply uniform computation to every input regardless of difficulty. We introduce TRIAGE, a tiered zero-shot framework that adaptively scales test-time compute by routing each audio sample through progressively richer reasoning stages: fast label-cosine scoring in a joint audio-text embedding space (Tier-L), structured matching with clinician-style descriptors (Tier-M), and retrieval-augmented large language model reasoning (Tier-H). A confidence-based router finalizes easy predictions early while allocating additional computation to ambiguous inputs, enabling nearly half of all samples to exit at the cheapest tier. Across nine respiratory classification tasks without task-specific training, TRIAGE achieves a mean AUROC of 0.744, outperforming prior zero-shot methods and matching or exceeding supervised baselines on multiple tasks. Our analysis show that test-time scaling concentrates gains where they matter: uncertain cases see up to 19% relative improvement while confident predictions remain unchanged at minimal cost.

Figures & tables

Figure 1: Tiered test-time inference improves zero-shot respiratory audio classification. Points show mean AUROC (%) averaged over 9 tasks versus budget tier. CLAP and AcuLa are zero-shot baselines (horizontal lines). TRIAGE (Fixed) evaluates three compute budgets by forcing all samples to run up to Tier-L/Tier-M/Tier-H (Low/Mid/High), while TRIAGE (Adaptive) reports the single operating point produced by confidence-based routing across tiers.

Figure 2: Overview of TRIAGE. A frozen audio–text encoder embeds recordings and medical text. Tier-L performs label cosine scoring with margin-based early exit; Tier-M matches clinical descriptors with rule voting; Tier-H retrieves nearest-neighbor reports and queries an off-the-shelf LLM, e.g., Gemini or GPT.

	TL-Finalized (High Conf.)			TM-Finalized (Medium Conf.)				TH-Escalated (Low Conf.)
Task	%	TL	Adapt.	%	TL	Adapt.	Rel. $\uparrow$	%	TL	Adapt.	Rel. $\uparrow$
UKCOV-EX-1	47	.645	.645	38	.562	.690	23%	15	.520	.705	36%
UKCOV-CO-1	62	.680	.680	25	.602	.662	10%	13	.566	.658	16%
CVID-CO-1	52	.779	.779	31	.711	.804	13%	17	.660	.783	19%
CVID-CO-2	43	.707	.707	37	.640	.695	9%	20	.638	.678	6%
ICBHI-LS-1	47	.761	.761	36	.679	.797	17%	17	.633	.802	27%

Table 2: AUROC (

\uparrow

) on nine respiratory audio classification tasks. We compare supervised linear-probing baselines (frozen encoders with task-specific training) and zero-shot methods (no task-specific training). CLAP (ZS) predicts via audio–text similarity with its own pretrained encoder. TRIAGE uses a frozen AcuLa audio–text encoder and allocates test-time compute across three tiers: Tier-L (label-name cosine scoring), Tier-M (clinician-approved descriptor template matching with rule voting), and Tier-H (FAISS

k

NN report retrieval + LLM reasoning). Adaptive routes examples hierarchically based on confidence.

\Delta

rows report gains of Adaptive over CLAP (ZS) and AcuLa (ZS).

Task ID	Classification Task	Baseline	20% Masking		50% Masking		Sensitivity
		AUROC	AUROC	$\Delta$	AUROC	$\Delta$	Rank
ICBHI-LS-1	COPD (Lung sounds)	0.832	0.794	-0.038	0.739	-0.093	High
UKCOV-EX-1	COVID-19 (Exhalation)	0.690	0.656	-0.034	0.607	-0.083	High
RESPTR-LS-1	COPD Severity (Lung)	0.698	0.662	-0.036	0.660	-0.038	High
CVID-CO-1	COVID-19 (Cough)	0.780	0.763	-0.017	0.734	-0.046	Medium
KAUH-LS-1	Obstructive Disease (Lung)	0.721	0.703	-0.018	0.690	-0.031	Medium

Table 3: Impact of random descriptor masking on Tier-M retrieval performance. Each task evaluated at three masking levels: 0% (baseline), 20%, and 50% random descriptor removal. Performance degradation (

\Delta

) color-coded: minimal (

|\Delta|<0.02

), moderate (

0.02\leq|\Delta|<0.05

), severe (

|\Delta|\geq 0.05

). Results based on 5 repeated maskings per rate.

Task ID	Classification	$d{=}1$	$d{=}3$	$d{=}5$	$d{=}8$	Optimal	Gain
	Task	(baseline)				$d$	$\Delta$
High Context Sensitivity (Gain $>$ 0.035)
RESPTR-LS-1	COPD Severity (Lung)	0.667	0.705	0.700	0.698	3	+0.038
CVID-CO-2	Gender (Cough)	0.646	0.682	0.682	0.680	3/5	+0.036
ICBHI-LS-1	COPD (Lung sounds)	0.785	0.812	0.818	0.817	5	+0.033
Medium Context Sensitivity (0.020 $<$ Gain $\leq$ 0.035)

Table 4: Tier-H LLM performance versus retrieval context depth (single call,

b{=}1

). Each task evaluated with 1, 3, 5, and 8 retrieved documents. Optimal context size (highest AUROC) highlighted in bold. Gain metrics show improvement from minimal context (

d{=}1

) to optimal depth. Results demonstrate diminishing returns beyond

d{=}3

for most tasks.

ID	Task	$\tau_{1}{=}0.30$				$\tau_{1}{=}0.45$				$\tau_{1}{=}0.60$
ID	Task	AUROC	%T-L	%T-M	%T-H	AUROC	%T-L	%T-M	%T-H	AUROC	%T-L	%T-M	%T-H
UKCOV-EX-1	COVID (Exh.)	0.708	41	43	16	0.713	34	48	18	0.711	27	54	19
UKCOV-CO-1	COVID (Cgh.)	0.675	57	29	14	0.678	47	37	16	0.677	39	43	18
CVID-CO-1	COVID (Cgh.)	0.817	46	36	18	0.819	38	42	20	0.818	30	48	22
CVID-CO-2	Gender (Cgh.)	0.702	36	43	21	0.705	31	47	22	0.703	23	54	23
ICBHI-LS-1	COPD (Lung)	0.844	43	39	18	0.844	33	47	20	0.846	27	52	21

Table 5: Effect of Tier-L confidence cutoff

\tau_{1}

on TRIAGE performance and compute. AUROC and tier usage (% of samples handled by Tier-L / Tier-M / Tier-H) for

\tau_{1}\in\{0.30,0.45,0.60\}

. The Tier-M

\rightarrow

Tier-H escalation threshold

\tau_{2}

is fixed across all settings, and Tier-H uses the same retrieval+LLM backend throughout.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Descriptor group	System option set	COPD prototype	Healthy prototype
Breath sound character	normal vesicular breathing; diminished or distant breath sounds; bronchial breathing with high pitch; absent breath sounds; amphoric hollow breathing; cavernous breathing sounds; harsh breathing with increased intensity	diminished or distant breath sounds	normal vesicular breathing
Wheeze presence	no wheeze detected; mild expiratory wheeze; moderate expiratory wheeze; severe expiratory wheeze; inspiratory wheeze present; biphasic wheeze (inspiratory and expiratory); polyphonic multiple wheeze; monophonic single wheeze	moderate expiratory wheeze	no wheeze detected
Respiratory phase timing	normal inspiratory to expiratory ratio 1:2; prolonged expiratory phase ratio 1:3 or greater; shortened expiratory phase ratio 1:1; prolonged inspiratory phase; equal inspiratory and expiratory phases; irregular variable phase timing; rapid shallow breathing pattern	prolonged expiratory phase ratio 1:3 or greater	normal inspiratory to expiratory ratio 1:2
Crackle characteristics	no crackles present; fine high-pitched crackles early inspiratory; fine crackles late inspiratory; coarse low-pitched crackles early inspiratory; coarse crackles throughout inspiration; bibasilar crackles at lung bases; diffuse crackles throughout lung fields; velcro-like crackles	coarse low-pitched crackles early inspiratory	no crackles present
Respiratory effort	normal and effortless; mildly increased effort; moderately labored; severely labored with accessory muscle use; paradoxical breathing pattern; shallow with minimal effort; gasping or air hunger pattern	moderately labored	normal and effortless
Spectral frequency profile	normal frequency distribution 100–1000 Hz; low frequency dominance below 400 Hz; high frequency dominance above 800 Hz; broadband frequency distribution; narrow band frequency concentration; bimodal frequency peaks; irregular frequency scatter	low frequency dominance below 400 Hz	normal frequency distribution 100–1000 Hz

Appendix

Table 6: Tier-2 descriptor groups and prototypical configurations for COPD vs Healthy lung sounds in task ICBHI-LS-1. For each descriptor group we list the full option set used in our system, and the specific option selected by the COPD and Healthy prototypes in task_rules .

Task ID	Classification Task	Audio Type	Samples	Class Distribution
COVID-19 Detection
UKCOV-EX-1	Covid vs. Non-covid	Exhalation	2,500	840 / 1,660
UKCOV-CO-1	Covid vs. Non-covid	Cough	2,500	840 / 1,660
CVID-CO-1	Covid vs. Non-covid	Cough	6,175	547 / 5,628
Demographic Classification
CVID-CO-2	Female vs. Male	Cough	7,263	2,468 / 4,795

Appendix

Table 7: Downstream evaluation tasks for respiratory audio classification. All tasks use binary or multi-class classification on audio recordings from real-world clinical and crowdsourced datasets.

ID	OCT	OCE	TL-only	TM-only	TH-only	Adaptive
UKCOV-EX-1	0.586	0.551	0.593	0.690	0.707	0.703
UKCOV-CO-1	0.701	0.629	0.627	0.652	0.670	0.672
CVID-CO-1	0.578	0.566	0.722	0.780	0.802	0.810
CVID-CO-2	0.795	0.721	0.668	0.640	0.682	0.700
ICBHI-LS-1	0.855	0.872	0.706	0.832	0.812	0.835
COSW-CO-1	0.685	0.674	0.717	0.695	0.700	0.728

Appendix

Table 9: Appendix: AUROC (

\uparrow

) for the omitted linear-probing baselines OPERA-CT and OPERA-CE, alongside our three-tier inference policies (same splits as Table 1 ).

Figure 3: Top-3 clinical reports retrieved for auscultation clips (Tier-H). Left: query spectrogram and its reference (or generated) report. Right: the three closest reports returned by FAISS in the shared embedding space.

Explore similar work

Zero-Shot Respiratory Sound Classification through LLM-Augmented Audio-Text Alignment
Aug 30, 2026Mustafa Talha İlerisoy, Hung Manh Pham, Mathias Funk +2Respiratory Sound Classification Audio-Language Models
Eindhoven University of Technology, Eindhoven, Netherlands · Singapore Management University, Singapore
Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis
Feb 16, 2026Pengfei Zhang, Tianxin Xie, Minghao Yang +1Respiratory Sound Classification Audio Generation
The Hong Kong University of Science and Technology (Guangzhou)
RespiraMFM: A Multimodal Foundation Model with Contrastive Audio-Language Alignment for Respiratory Disease Identification
Jun 8, 2026Shakhrul Iman Siam, Tiantian Feng, Jiankun Zhang +2Healthcare Audio-Language Models
The Ohio State University · University of Southern California · University of Chicago

Aug 30, 2026J/K move · Enter open · S save

Zero-Shot Respiratory Sound Classification through LLM-Augmented Audio-Text Alignment

Mustafa Talha İlerisoy, Hung Manh Pham, Mathias Funk, Mykola Pechenizkiy+1

Eindhoven University of Technology, Eindhoven, Netherlands · Singapore Management University, Singapore

Respiratory Sound Classification Audio-Language Models Multimodal Contrastive Learning

Open paper AskPDF arXiv

Self-supervised respiratory encoders lack semantic grounding in clinical domain needed for zero-shot inference, limiting their utility without task-specific labeled data. We propose a framework that aligns these encoders with medical terminology in a shared latent space turning them into a zero-shot-capable foundation model. To address paired data scarcity, we use a medical LLM to synthesize structured reports from metadata, creating dense semantic anchors for contrastive learning. Our training combines a sigmoid-based contrastive loss with encoder's native SSL objective and similarity-aware negative sampling to sharpen pathological boundaries. Across 9 tasks on 6 datasets, our method achieves a 61.3% mean zero-shot AUC, surpassing CLAP (51.4%) and Qwen2-Audio (54.9%) while reaching the highest linear probing AUC (71.6%) with only 43% of data used by full-scale baselines, showing that structured semantic alignment outperforms large-scale, general-purpose models in clinical diagnostics.

Abstract

Explore similar work

Zero-Shot Respiratory Sound Classification through LLM-Augmented Audio-Text Alignment

Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis

RespiraMFM: A Multimodal Foundation Model with Contrastive Audio-Language Alignment for Respiratory Disease Identification