Paper deep dive
RAST: Resolution-Aware Privileged Structure Transfer for Low-Resolution Audio Activity Recognition
Ji Hwan Park, Gautham Krishna Gudur, Yufei Shen, Dawei Liang, Edison Thomaz
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 10/4/2026, 3:02:32 AM
Summary
The paper introduces RAST (Resolution-Aware Privileged Structure Transfer), a framework for low-resolution audio activity recognition. It addresses the performance degradation caused by reduced sampling rates by using high-resolution audio as privileged information during training. RAST compresses high-resolution teacher representations into discrete tokens and neighborhood structures, which are then used to warm-start and align a low-resolution student model. Experiments on SAMoSA and AudioIMU datasets demonstrate that RAST significantly outperforms baselines like LR-only training and standard knowledge distillation.
Entities (9)
Relation Signals (8)
RAST → evaluatedon → SAMoSA
confidence 95% · Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms...
RAST → evaluatedon → AudioIMU
confidence 95% · Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms...
RAST → solves → Human Activity Recognition
confidence 95% · RAST: Resolution-Aware Privileged Structure Transfer for Low-Resolution Audio Activity Recognition
RAST → targets → Low-Resolution Audio
confidence 90% · RAST... for Low-Resolution Audio Activity Recognition... requiring only LR audio at inference
RAST → uses → High-Resolution Audio
confidence 90% · RAST... compresses HR teacher representations... HR audio is available during training
RAST → outperforms → Knowledge Distillation
confidence 85% · RAST consistently outperforms LR-only training and direct teacher-transfer baselines... Compared with KD... these gains indicate...
RAST → usestechnique → Hard-negative mining
confidence 85% · refined through local hard-negative-mining (HNM) alignment
RAST → usesarchitecture → PANNs
confidence 80% · All methods use the same PANNs-based audio encoder
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Audio is increasingly used for human activity recognition (HAR) because it captures object interactions, environmental events, and contextual cues in everyday environments. High-resolution (HR) audio provides rich acoustic information for model development but incurs substantial energy and storage costs and may expose sensitive speech content. Low-resolution (LR) audio offers a more privacy-preserving and resource-efficient alternative for deployment, but reduced sampling rates can remove acoustic cues essential for activity recognition, leading to significant performance degradation. We formulate this training-deployment mismatch as sensor-resolution privileged learning, in which HR audio is available during training, while inference relies exclusively on LR audio. We propose RAST, a resolution-aware transfer framework that compresses HR teacher representations by preserving token-level information and neighborhood structure before performing localized HR-LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR-only training and direct teacher-transfer baselines, improving LR-only recognition by up to approximately 7.8% while requiring only LR audio at inference.
Tags
Links
- Source: https://arxiv.org/abs/2609.38780v1
- Canonical: https://arxiv.org/abs/2609.38780v1
Trouble viewing inline? Open PDF directly →
Full Text
32,786 characters extracted from source content.
Expand or collapse full text
RAST: Resolution-Aware Privileged Structure Transfer for Low-Resolution Audio Activity Recognition Conference: The 4th International Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems; October 26–30, 2026; Austin, TX, USAThe 4th International Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems (HumanSys ’26), October 26–30, 2026, Austin, TX, USADOI: 10.1145/3842436.3843818ISBN: 979-8-4007-2959-1/2026/10CCS: Human-centered computingCCS: Computing methodologies Machine learningCCS: Computing methodologies Learning paradigms Ji Hwan Park Affiliation: The University of Texas at Austin, Austin, Texas, USA email: jihwanpark98@utexas.edu , Gautham Krishna Gudur Affiliation: The University of Texas at Austin, Austin, Texas, USA email: gauthamkrishna@utexas.edu , Yufei Shen Affiliation: The University of Texas at Austin, Austin, Texas, USA email: shenyufei@utexas.edu , Dawei Liang Affiliation: The University of Texas at Austin, Austin, Texas, USA email: dawei.liang@utexas.edu and Edison Thomaz Affiliation: The University of Texas at Austin, Austin, Texas, USA email: ethomaz@utexas.edu © c Abstract. Audio is increasingly used for human activity recognition (HAR) because it captures object interactions, environmental events, and contextual cues in everyday environments. High-resolution (HR) audio provides rich acoustic information for model development but incurs substantial energy and storage costs and may expose sensitive speech content. Low-resolution (LR) audio offers a more privacy-preserving and resource-efficient alternative for deployment, but reduced sampling rates can remove acoustic cues essential for activity recognition, leading to significant performance degradation. We formulate this training–deployment mismatch as sensor-resolution privileged learning, in which HR audio is available during training, while inference relies exclusively on LR audio. We propose RAST, a resolution-aware transfer framework that compresses HR teacher representations by preserving token-level information and neighborhood structure before performing localized HR–LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR-only training and direct teacher-transfer baselines, improving LR-only recognition by up to ∼7.8% 7.8\% while requiring only LR audio at inference. Keywords: human-centered sensing, audio sensing, privileged learning †c-license: by 1. Introduction Audio is an attractive modality for human-centered sensing because everyday sounds encode rich contextual information about human behavior, object interactions, environmental context, and so on. These capabilities have motivated acoustic sensing for activities of daily living, classification from low-sampled microphone input, and wearable HAR (Pires et al., 2017; Liang and Thomaz, 2019; Liang et al., 2024). However, continuous acoustic sensing imposes practical and social constraints, including substantial bandwidth, storage, computation, and energy costs (Georgiev et al., 2014), while always-on microphones raise concerns about privacy, trust, and unintended speech capture (Chhaglani et al., 2025). Recent work has therefore explored restricted-bandwidth and low-resolution audio as privacy-aware sensing alternatives, including analyses of speech intelligibility under bandwidth reduction (Liu et al., 2024). Audio-based HAR thus presents a broader human-centered systems challenge: reliably recognizing meaningful behavior while limiting sensing cost, sensitive-data exposure, and resource consumption (Lu et al., 2009; Consolvo et al., 2008). Low-resolution (LR) audio is well suited for resource-constrained and privacy-sensitive deployment, but reducing the sampling rate can remove acoustic cues critical for distinguishing activities. High-resolution (HR) audio preserves high-frequency and transient structure associated with activities such as chopping, brushing, clapping, pouring, and appliance use whereas a deployed model may observe only the LR signal. Figure 1 illustrates this resolution gap: LR audio suppresses salient acoustic structure, and held-out recognition accuracy decreases as the sampling rate diverges from the 16-kHz HR reference. Prior work has shown that large-scale pretrained audio representations can improve downstream sound and activity recognition (Aytar et al., 2016; Kong et al., 2020; Gudur et al., 2026), while recent work has explored self-distillation for low-sampled microphone input (Liang et al., 2024). Motivated by this gap, we study a privileged-learning setting in which HR audio is available only during training, while inference operates exclusively on LR audio. The objective is to transfer discriminative HR information without requiring HR sensing at deployment. Figure 1. Resolution gap in audio-based activity recognition. (a) High-resolution (HR) and Low-resolution (LR) views of the same audio window. (b) LR accuracy decreases as the sampling rate is reduced relative to the 16-kHz HR reference. Resolution gap between High-resolution (HR) and Low-resolution (LR) in audio-based activity recognition. This setting extends beyond conventional teacher-student transfer. Classical knowledge distillation typically compresses a larger or more capable teacher into a smaller student given comparable inputs (Hinton et al., 2015). Generalized distillation and learning using privileged information allow access to additional training-only information (Vapnik and Vashist, 2009; Lopez-Paz et al., 2016), while cross-modal distillation transfers supervision across paired modalities (Gupta et al., 2016; Aytar et al., 2016). In our setting, the primary challenge is instead limited input observability: the LR student receives a systematically degraded view of the same acoustic event. Consequently, some evidence encoded by the HR teacher may be absent from the LR signal, making direct imitation of the full HR representation both restrictive and potentially infeasible. To address this challenge, we propose Resolution-Aware Privileged Structure Transfer (RAST), a structured HR-to-LR transfer framework tailored to sensor-resolution mismatch. Rather than forcing the LR student to reproduce the complete HR embedding space, RAST compresses teacher representations into intermediate structure comprising HR-derived tokens that summarize coarse teacher states and teacher neighborhoods that encode local geometry. The LR student is first warm-started using this privileged structure and then refined through local hard-negative-mining (HNM) alignment. This design builds on prior work showing that intermediate hints, relational geometry, contrastive objectives, and information bottlenecks can transfer knowledge beyond final logits (Romero et al., 2015; Park et al., 2019; Tian et al., 2020; Tishby et al., 2000). RAST therefore focuses transfer on HR-derived structure that is both discriminative and recoverable from LR audio, avoiding imitation of fine-grained information that may be unobservable after resolution reduction. This paper makes three contributions. (1) We formulate lower-resolution audio HAR as a sensor-resolution privileged learning problem, where an LR student must learn from an HR teacher despite limited input observability. (2) We design RAST, a resolution-aware HR-to-LR transfer framework that combines HR-derived token bottlenecks, teacher-neighborhood initialization, and local representation alignment. (3) We evaluate RAST on two audio-HAR datasets – SAMoSA (Mollyn et al., 2022) and AudioIMU (Liang et al., 2022), under subject-independent protocols. Relative to LR-only training, RAST improves balanced accuracy and macro-F1 by 6.7%6.7\% and 7.5%7.5\% on SAMoSA, and by 6.1%6.1\% and 7.8%7.8\% on AudioIMU respectively. Together, these results demonstrate that privileged HR audio available only during training can substantially improve LR activity recognition while preserving an exclusively low-resolution sensing path at inference. 2. Related Work Knowledge transfer with privileged information. Knowledge distillation transfers teacher knowledge through softened outputs (Hinton et al., 2015), with subsequent work extending transfer to intermediate features, contrastive objectives, and relational structure (Romero et al., 2015; Tian et al., 2020; Park et al., 2019). Learning using privileged information allows additional training-only information (Vapnik and Vashist, 2009), while generalized distillation integrates such information into teacher–student transfer (Lopez-Paz et al., 2016). Related paradigms include dataset distillation for compact audio-classification training sets (Gudur and Thomaz, 2024), zero-shot distillation for statistical heterogeneity in privacy-preserving federated learning (Gudur and Perepu, 2021), and cross-modal distillation across paired modalities (Gupta et al., 2016; Aytar et al., 2016). HR-to-LR audio transfer introduces a distinct observability constraint because the student receives a lower-resolution view of the same event. Directly matching HR predictions or representations may therefore impose ill-suited targets that are partially unobservable or unrecoverable from LR input. Deployment-aware activity sensing. Human-centered sensing must balance recognition accuracy with sensing cost, privacy, participant burden, and robustness. Mobile acoustic sensing has enabled context and activity recognition (Liang and Thomaz, 2019), while other prior work has characterized deployment constraints, user variability, and sensor heterogeneity (Consolvo et al., 2008; Wang et al., 2019; Stisen et al., 2015; Sundaramoorthy et al., 2018). Privacy and trust concerns around always-on audio have further motivated low-resolution and bandwidth-restricted sensing (Chhaglani et al., 2025), including self-distillation for 1–2 kHz microphone inputs (Liang et al., 2024) and discussions on privacy–utility tradeoffs in low-frequency speech (Liu et al., 2024). RAST leverages HR audio only as privileged training supervision while requiring only LR audio at inference. Compact teacher structure. Information-bottleneck methods seek compact representations that preserve task-relevant information while suppressing unnecessary detail (Tishby et al., 2000; Alemi et al., 2017), while discrete representation learning provides mechanisms for constructing compact latent codes (van den Oord et al., 2017). More broadly, hint-based, relational, and contrastive distillation demonstrate that intermediate structure can convey useful teacher knowledge beyond hard labels or final logits (Romero et al., 2015; Park et al., 2019; Tian et al., 2020). RAST adapts this principle to sensor-resolution mismatch by transferring compact HR-derived tokens and local teacher-neighborhood structure instead of requiring the LR student to reproduce the full HR embedding space. Table 1. Subject-held-out audio activity recognition performance. BA denotes balanced accuracy. The HR teacher is shown as the high-resolution reference, while the remaining methods use LR input at inference. Parentheses report relative gains over the LR-only baseline. SAMoSA AudioIMU Method BA Macro-F1 BA Macro-F1 HR teacher 0.8352±0.00830.8352± 0.0083 0.8307±0.00750.8307± 0.0075 0.7246±0.00170.7246± 0.0017 0.7066±0.00320.7066± 0.0032 LR-only 0.6123±0.00540.6123± 0.0054 0.6030±0.00460.6030± 0.0046 0.5157±0.00330.5157± 0.0033 0.4933±0.00310.4933± 0.0031 KD 0.6109±0.00350.6109± 0.0035 0.6053±0.00470.6053± 0.0047 0.5124±0.00540.5124± 0.0054 0.4843±0.00640.4843± 0.0064 InfoNCE 0.6137±0.00410.6137± 0.0041 0.6086±0.00160.6086± 0.0016 0.5161±0.00380.5161± 0.0038 0.4994±0.00400.4994± 0.0040 ProtoNCE 0.6112±0.00650.6112± 0.0065 0.6055±0.00790.6055± 0.0079 0.5154±0.00230.5154± 0.0023 0.4921±0.00360.4921± 0.0036 HNM 0.6218±0.00820.6218± 0.0082 0.6093±0.00860.6093± 0.0086 0.5130±0.00190.5130± 0.0019 0.5079±0.00320.5079± 0.0032 RAST 0.6532±0.0072(+6.7%)0.6532± 0.0072 (+6.7\%) 0.6482±0.0067(+7.5%)0.6482± 0.0067 (+7.5\%) 0.5474±0.0042(+6.1%)0.5474± 0.0042 (+6.1\%) 0.5320±0.0055(+7.8%)0.5320± 0.0055 (+7.8\%) 3. Methodology 3.1. Problem Formulation Let =(xiH,xiL,yi)i=1ND=\(x_i^H,x_i^L,y_i)\_i=1^N denote the training set, where xiHx_i^H and xiLx_i^L are paired high-resolution (HR) and low-resolution (LR) views of the same audio segment, respectively, and yi∈1,…,Cy_i∈\1,…,C\ is the activity label. The HR view is available only during training. At inference time, the deployed model observes only xLx^L and must predict its activity label. This formulation separates the deployment constraint from the model-development setting: richer HR audio can be collected or retained for training, while the deployed classifier remains an LR-only system. We train an HR teacher THT_H and an LR student SLS_L, each comprising an encoder and a classifier head: (1) ziH=eH(xiH),piH=gH(ziH),z_i^H=e_H(x_i^H), p_i^H=g_H(z_i^H), (2) ziL=eL(xiL),piL=gL(ziL).z_i^L=e_L(x_i^L), p_i^L=g_L(z_i^L). where ziHz_i^H and ziLz_i^L denote latent representations. The standard LR baseline minimizes the cross-entropy objective, (3) ℒCE=1N∑iCE(piL,yi).L_CE= 1N _iCE(p_i^L,y_i). RAST augments this LR classification objective with compact structural supervision extracted from the HR teacher. Figure 2. RAST training and deployment pipeline. HR audio provides token- and neighborhood-level privileged supervision only during training, while deployment uses only the LR student and LR audio.Overall RAST training and deployment pipeline with three stages. 3.2. Framework Overview RAST comprises three stages. In Stage 1, the HR teacher is trained for activity recognition, and its intermediate representations are distilled into two forms of compact privileged structure: discrete HR-derived tokens and local teacher neighborhoods. In Stage 2, the LR student uses these structures for initialization (warm start) and is subsequently refined through local HR–LR representation alignment. In Stage 3, the HR teacher and all auxiliary transfer targets are discarded. Deployment therefore requires only the LR student and LR audio input. Figure 2 summarizes the privileged train-time transfer path and the LR-only inference path. RAST transfers compact, structured HR knowledge rather than requiring the LR student to reproduce the full HR embedding space. Because downsampling removes acoustic evidence available to the teacher, direct HR–LR representation matching can impose partially unrecoverable targets. RAST therefore transfers bottlenecked HR-derived structure followed by local hard-negative alignment. 3.3. Privileged Structure Extraction HR-derived tokens. After training the HR teacher THT_H, we extract representations from a fixed intermediate block and cluster the resulting feature space using K-means. Each training example receives a token label: (4) ci=argmink∈1,…,K‖z¯iH−μk‖22,c_i= _k∈\1,…,K\ \| z_i^H- _k \|_2^2, where z¯iH z_i^H denotes the normalized HR teacher representation and μk _k the kkth cluster centroid. These tokens provide a compact discretization of the HR representation space, exposing coarse teacher structure to the LR student. They serve only as auxiliary training targets and are discarded at inference. Teacher neighborhoods. Although HR-derived tokens capture coarse regions of the teacher representation space, they do not preserve its local geometry. We therefore construct teacher-neighborhood distributions within each mini-batch. For samples i and j, the HR pairwise similarity is: (5) sijH=sim(z¯iH,z¯jH),s^H_ij=sim( z_i^H, z_j^H), where simsim denotes cosine similarity. The neighborhood distribution induced by the HR teacher for anchor i is: (6) qijH=exp(sijH/τ)∑m≠iexp(simH/τ),j≠i.q^H_ij= (s^H_ij/τ) _m≠ i (s^H_im/τ), j≠ i. where τ is the neighborhood temperature. The LR student induces an analogous distribution qijLq^L_ij from zLz^L. We match these distributions with: (7) ℒnbr=1B∑i=1BKL(qiH∥qiL),L_nbr= 1B _i=1^BKL(q_i^H\|q_i^L), where B is the mini-batch size. 3.4. Student Optimization and Alignment The LR student is first trained with label, token, and teacher-neighborhood supervision: (8) ℒwarm=ℒCE+λtokCE(riL,ci)+βℒnbr,L_warm=L_CE+ _tokCE(r_i^L,c_i)+ _nbr, where riLr_i^L is the LR token prediction head. This warm start shapes the LR representation before direct HR–LR alignment. We then apply local hard-negative-mining (HNM) alignment between LR and HR teacher embeddings. For anchor i, (ziL,ziH)(z_i^L,z_i^H) forms the positive pair, while k(i)N_k(i) contains the top-k most similar non-matching HR examples selected as hard negatives in the mini-batch. We optimize: (9) ℒHNM=−1B∑ilogexp(sim(ziL,ziH)/γ)∑j∈k(i)exp(sim(ziL,zjH)/γ).L_HNM=- 1B _i (sim(z_i^L,z_i^H)/γ) _j _k(i) (sim(z_i^L,z_j^H)/γ). HNM focuses alignment on locally confusable examples rather than matching the full HR manifold uniformly. The structured warm start further improves alignment by first exposing recoverable HR-derived tokens and neighborhood structure to the LR student. 4. Experiments 4.1. Datasets and Sensing Setup We evaluate our proposed framework on wearable-based HAR datasets collected in home and semi-naturalistic environments: AudioIMU (Liang et al., 2022) and SAMoSA (Mollyn et al., 2022). AudioIMU contains 15 participants performing 23 activities of daily living, recorded with a Fossil Gen 4 smartwatch at 22.05 kHz audio and 50 Hz IMU. SAMoSA contains 20 participants recorded across diverse real-world environments using a Fossil Gen 5 smartwatch at 16 kHz audio and 50 Hz IMU. Both datasets provide synchronized acoustic and inertial streams; we use only audio. Audio is segmented into aligned 5-second windows. The HR branch uses 16 kHz audio, while the LR branch simulates constrained sensing by reducing the effective sampling rate to 1.25 kHz before computing the same log-mel representation. HR and LR inputs are paired views of the same activity window, with only LR audio available at test time. 4.2. Evaluation Protocol We use subject-independent evaluation to measure generalization to unseen participants. SAMoSA uses leave-one-group-out validation over six participant groups, while AudioIMU uses leave-one-participant-out validation over participants 01–15. Results are reported as mean ± standard deviation over five seeds. We report balanced accuracy (BA) and macro-F1; BA accounts for uneven class frequencies, while macro-F1 measures per-class recognition without allowing frequent classes to dominate. 4.3. Methods and Implementation All methods use the same PANNs-based audio encoder (Kong et al., 2020), pretrained on AudioSet (Gemmeke et al., 2017), and identical subject-held-out folds. Baselines isolate different forms of HR-to-LR knowledge transfer: LR-only uses no privileged supervision; Knowledge distillation (KD) transfers teacher logits (Hinton et al., 2015); InfoNCE performs instance-level contrastive alignment (Oord et al., 2018; Tian et al., 2020); Prototype NCE (ProtoNCE) aligns class-level prototypes inspired by prototypical and prototype-contrastive representation learning (Snell et al., 2017; Li et al., 2021); and hard-negative mining (HNM) directly aligns local HR–LR representations using hard-negative selection (Robinson et al., 2021). RAST augments HNM with the proposed token and teacher-neighborhood warm start, evaluating whether compact HR structure transfers more effectively than direct teacher imitation under resolution mismatch. The HR token codebook uses K=100K=100 clusters extracted from an intermediate teacher block. Unless otherwise stated, RAST uses λtok=1.0 _tok=1.0, β=0.1β=0.1, neighborhood temperature τ=0.2τ=0.2, and HNM temperature γ=0.2γ=0.2. Experiments were conducted on a workstation with eight NVIDIA RTX A4500 GPUs, each with 20 GB of memory. 5. Results 5.1. Discussion on Overall Results Table 1 compares RAST with LR-only training and direct HR-to-LR transfer baselines. RAST consistently achieves the best LR inference performance on both datasets. On SAMoSA, balanced accuracy improves from 0.61230.6123 to 0.65320.6532 and macro-F1 from 0.6030.603 to 0.64820.6482, corresponding to relative gains of 6.7%6.7\% and 7.5%7.5\%, respectively. On AudioIMU, balanced accuracy increases from 0.51570.5157 to 0.54740.5474 and macro-F1 from 0.49330.4933 to 0.5320.532, yielding relative gains of 6.1%6.1\% and 7.8%7.8\% respectively. Compared with KD, InfoNCE, ProtoNCE, and HNM, these gains indicate that HR supervision using teacher is more effective when first distilled into compact tokens and local neighborhood structure before localized HR–LR alignment. 5.2. Ablation Study Table 2. Component-level performance of RAST on SAMoSA before and after HNM alignment. Balanced Accuracy (BA) and macro-F1 are reported. Warm start After HNM Init. BA F1 BA F1 CE 0.61230.6123 0.60300.6030 0.62180.6218 0.60930.6093 Token 0.62180.6218 0.61050.6105 0.64430.6443 0.62400.6240 Nbr. 0.62430.6243 0.61310.6131 0.64420.6442 0.62250.6225 Token+Nbr. 0.63080.6308 0.61820.6182 0.65320.6532 0.64820.6482 Table 2 isolates how structured warm start supervision organizes the LR representation space before subsequent HNM alignment. With CE-only initialization, HNM improves balanced accuracy from 0.61230.6123 to 0.62180.6218. Adding HR-token supervision raises the warm-start score to 0.62180.6218 and the post-HNM result to 0.64430.6443, indicating that the token bottleneck provides coarse, activity-relevant anchors that facilitate subsequent local alignment. Teacher-neighborhood regularization further improves the alignment interface. Although neighborhood supervision alone yields a moderate warm-start score, its post-HNM performance is comparable to token supervision, suggesting that local teacher geometry helps organize nearby examples before hard-negative alignment. Combining tokens and neighborhoods yields the strongest warm start (0.63080.6308 BA, 0.61820.6182 macro-F1) and the final result (0.65320.6532 BA, 0.64820.6482 macro-F1). These results support the RAST design: tokenization supervision provides compact HR-derived learnable anchors, neighborhood regularization structures their local geometry around these anchors, and HNM then aligns the resulting LR representation more effectively. 5.3. Embedding Structure Analysis Three t-SNE visualizations of intermediate embeddings for high-resolution, low-resolution, and RAST. Figure 3. t-SNE visualization of intermediate embeddings. RAST transfers HR-derived structure to the LR model, producing representations that are better aligned with activity structure than the LR baseline.Three t-SNE visualizations of intermediate embeddings for high-resolution, low-resolution, and RAST. Figure 3 illustrates the HR–LR representation gap and the role of RAST’s structured warm start. The HR teacher preserves more coherent local neighborhood geometry, with related activities forming better-organized regions. In contrast, the LR-only model produces a more entangled representation, particularly for activities that depend on fine acoustic cues attenuated by downsampling. RAST restructures the transfer problem before local alignment. Rather than directly matching the full HR embedding space, it first exposes the LR student to compact HR-derived tokens and teacher-neighborhood structure, providing coarse anchors and local geometric constraints for subsequent HNM. Consistent with the ablation results in Table 2, the resulting RAST embedding representation in Figure 3 shows improved separation among locally confusable activities and aligns with the overall gains in Table 1. 6. Conclusion In this paper, we present RAST, a resolution-aware privileged structure transfer framework for low-resolution audio activity recognition. RAST uses HR audio only during training, compressing teacher representations into token- and neighborhood-level supervision before local HR–LR alignment. Across SAMoSA and AudioIMU datasets, RAST consistently outperforms LR-only training and direct teacher-student transfer under subject-independent evaluation. The HR–LR gap motivates methods that explicitly model resolution mismatch and LR observability. Future work includes student-aware teachers, adaptive privileged bottlenecks, and alignment objectives that preserve activity-relevant structure recoverable from LR audio, with broader evaluation of privacy and real-world deployment tradeoffs. References Alemi et al. (2017) A. A. Alemi, I. Fischer, J. V. Dillon, and K. Murphy Deep variational information bottleneck. In International Conference on Learning Representations, Cited by: §2. Aytar et al. (2016) Y. Aytar, C. Vondrick, and A. Torralba SoundNet: learning sound representations from unlabeled video. In Advances in Neural Information Processing Systems, Vol. 29, p. . Cited by: §1, §1, §2. Chhaglani et al. (2025) B. Chhaglani, S. S. Gomasta, Y. Agarwal, J. Gummeson, and P. Shenoy FeatureSense: protecting speaker attributes in always-on audio sensing system. arXiv preprint arXiv:2505.24115. Cited by: §1, §2. Consolvo et al. (2008) S. Consolvo, P. Klasnja, D. W. McDonald, D. Avrahami, J. Froehlich, L. LeGrand, R. Libby, K. Mosher, and J. A. Landay Activity sensing in the wild: a field trial of UbiFit Garden. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, p. 1797–1806. Cited by: §1, §2. Gemmeke et al. (2017) J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter Audio Set: an ontology and human-labeled dataset for audio events. In IEEE International Conference on Acoustics, Speech and Signal Processing, p. 776–780. Cited by: §4.3. Georgiev et al. (2014) P. Georgiev, N. D. Lane, K. K. Rachuri, and C. Mascolo DSP.ear: leveraging co-processor support for continuous audio sensing on smartphones. In Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems, p. 295–309. Cited by: §1. Gudur et al. (2026) G. K. Gudur, M. Malu, T. Khandait, R. R. Azghan, A. Rayas, P. Turaga, J. Ghosh, H. Ghasemzadeh, E. Thomaz, and G. Pedrielli PCL: partitioned continual learning via unsupervised latent experts for audio classification. ICML 2026 Workshop on Learning to Listen: Machine Learning for Audio. Cited by: §1. Gudur and Perepu (2021) G. K. Gudur and S. K. Perepu Zero-Shot Federated Learning with New Classes for Audio Classification. In Interspeech 2021, p. 1579–1583. Cited by: §2. Gudur and Thomaz (2024) G. K. Gudur and E. Thomaz Dataset distillation for audio classification: a data-efficient alternative to active learning. NeurIPS 2024 Efficient Natural Language and Speech Processing (ENLSP-IV) Workshop. Cited by: §2. Gupta et al. (2016) S. Gupta, J. Hoffman, and J. Malik Cross modal distillation for supervision transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 2827–2836. Cited by: §1, §2. Hinton et al. (2015) G. Hinton, O. Vinyals, and J. Dean Distilling the knowledge in a neural network. In NeurIPS Deep Learning and Representation Learning Workshop, Cited by: §1, §2, §4.3. Kong et al. (2020) Q. Kong, Y. Cao, T. Iqbal, Y. Wang, W. Wang, and M. D. Plumbley PANNs: large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing 28, p. 2880–2894. Cited by: §1, §4.3. Li et al. (2021) J. Li, P. Zhou, C. Xiong, and S. C. H. Hoi Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, Cited by: §4.3. Liang et al. (2022) D. Liang, G. Li, R. Adaimi, R. Marculescu, and E. Thomaz AudioIMU: enhancing inertial sensing-based activity recognition with acoustic models. In Proceedings of the 2022 ACM International Symposium on Wearable Computers, p. 44–48. Cited by: item (3), §4.1. Liang and Thomaz (2019) D. Liang and E. Thomaz Audio-based activities of daily living (adl) recognition with large-scale acoustic embeddings from online videos. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 3 (1). Cited by: §1, §2. Liang et al. (2024) D. Liang, A. Zhang, D. Harwath, and E. Thomaz Improving audio classification with low-sampled microphone input: an empirical study using model self-distillation. In Proceedings of Interspeech, Cited by: §1, §1, §2. Liu et al. (2024) A. Liu, P. Vunderink, J. Vargas Quiros, C. Raman, and H. Hung How Private is Low-Frequency Speech Audio in the Wild? An Analysis of Verbal Intelligibility by Humans and Machines. In Interspeech 2024, p. 2725–2729. Cited by: §1, §2. Lopez-Paz et al. (2016) D. Lopez-Paz, L. Bottou, B. Scholkopf, and V. Vapnik Unifying distillation and privileged information. In International Conference on Learning Representations, Cited by: §1, §2. Lu et al. (2009) H. Lu, W. Pan, N. D. Lane, T. Choudhury, and A. T. Campbell SoundSense: scalable sound sensing for people-centric applications on mobile phones. In Proceedings of the 7th International Conference on Mobile Systems, Applications, and Services, p. 165–178. Cited by: §1. Mollyn et al. (2022) V. Mollyn, K. Ahuja, D. Verma, C. Harrison, and M. Goel SAMoSA: sensing activities with motion and subsampled audio. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 6 (3), p. 1–19. Cited by: item (3), §4.1. Oord et al. (2018) A. v. d. Oord, Y. Li, and O. Vinyals Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748. Cited by: §4.3. Park et al. (2019) W. Park, D. Kim, Y. Lu, and M. Cho Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 3967–3976. Cited by: §1, §2, §2. Pires et al. (2017) I. M. Pires, N. M. Garcia, N. Pombo, and F. Florez-Revuelta User environment detection with acoustic sensors embedded on mobile devices for the recognition of activities of daily living. arXiv preprint arXiv:1711.00124. Cited by: §1. Robinson et al. (2021) J. Robinson, C. Chuang, S. Sra, and S. Jegelka Contrastive learning with hard negative samples. In International Conference on Learning Representations, Cited by: §4.3. Romero et al. (2015) A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio FitNets: hints for thin deep nets. In International Conference on Learning Representations, Cited by: §1, §2, §2. Snell et al. (2017) J. Snell, K. Swersky, and R. Zemel Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, Cited by: §4.3. Stisen et al. (2015) A. Stisen, H. Blunck, S. Bhattacharya, T. S. Prentow, M. B. Kjærgaard, A. Dey, T. Sonne, and M. M. Jensen Smart devices are different: assessing and mitigating mobile sensing heterogeneities for activity recognition. p. 127–140. Cited by: §2. Sundaramoorthy et al. (2018) P. Sundaramoorthy, G. K. Gudur, M. R. Moorthy, R. N. Bhandari, and V. Vijayaraghavan HARNet: towards on-device incremental learning using deep ensembles on constrained devices. In Proceedings of the 2nd International Workshop on Embedded and Mobile Deep Learning, p. 31–36. Cited by: §2. Tian et al. (2020) Y. Tian, D. Krishnan, and P. Isola Contrastive representation distillation. In International Conference on Learning Representations, Cited by: §1, §2, §2, §4.3. Tishby et al. (2000) N. Tishby, F. C. Pereira, and W. Bialek The information bottleneck method. arXiv preprint physics/0004057. Cited by: §1, §2. van den Oord et al. (2017) A. van den Oord, O. Vinyals, and K. Kavukcuoglu Neural discrete representation learning. In Advances in Neural Information Processing Systems, Cited by: §2. Vapnik and Vashist (2009) V. Vapnik and A. Vashist A new learning paradigm: learning using privileged information. Neural Networks 22 (5–6), p. 544–557. Cited by: §1, §2. Wang et al. (2019) J. Wang, Y. Chen, S. Hao, X. Peng, and L. Hu Deep learning for sensor-based activity recognition: a survey. Pattern Recognition Letters 119, p. 3–11. Cited by: §2.