PaperlayerView exact revision on arXiv

Source-linked AI summary

SincDPNet: Interpretable Raw-Waveform Bathroom Activity Recognition for Assistive Living

Debolina Chowdhury, Suman Samui, Sujoy Saha

arXiv:2609.34907v1cs.SDcs.AIcs.LG

TL;DR

Bathroom acoustic recognition could support assistive monitoring where continuous video is undesirable, but practical models need to be compact, interpretable, and evaluated across environments. This paper introduces a seven-class dataset and a compact, interpretable raw-waveform classifier, then reports performance under environment-disjoint evaluation. Its compact configuration achieves 75.7% accuracy with 2,848 parameters, while broader generalization and deployment testing remain future work.

  • Problem

    Bathroom activities relevant to assistive living are underrepresented in public acoustic datasets, while recognition also calls for interpretable and resource-efficient models evaluated across recording environments.

    Source: §1 · p. 1, §2.5 · p. 3

  • Method

    The paper introduces SnaanGhar7 and SincDPNet, pairing a two-parameter-per-filter learnable sinc front end with a depthwise-separable waveform model evaluated using environment-disjoint splits.

    Source: §1 · p. 1, §2.5 · p. 3, §3 · p. 4

  • Results

    75.7% accuracy, 0.661 macro-F1, and 0.716 MCC are achieved by the 2,848-parameter N_f=25 SincDPNet on the held-out environment.

    Source: §10 · p. 26

  • Takeaways & Limitations

    The compact model’s interpretable frequency bands and environment-disjoint evaluation support examining acoustic structure and model errors in the studied bathroom-event setting.

    Source: §1 · p. 1, §2.5 · p. 3, §9 · p. 25

    Generalization across broader room, device, and acquisition conditions, and deployment measurements on microcontroller-class hardware, remain future evaluation needs.

    Source: §10 · p. 27

Abstract

from arXiv · show

Bathroom acoustic-event recognition can support ambient assisted living in settings where continuous video monitoring is undesirable. However, practical deployment requires models that are compact, interpretable, and robust to changes in the recording environment. This work introduces \dataset{}, a seven-class bathroom acoustic-event dataset containing 21{,}387 annotated clips recorded across five environments, and proposes SincDPNet, a compact raw-waveform classifier with a learnable sinc filter bank followed by a depthwise-separable convolutional body. Each sinc filter is controlled by two frequency parameters, allowing the learned passbands to be inspected directly in hertz while keeping the front end small. To reduce room-specific leakage, recording sessions and environments are separated before overlapping windows are assigned to the training, validation, and test partitions. We further use multi-objective Bayesian optimization as a design tool to examine the validation performance--model-size trade-off across 24 configurations. The selected designs span different operating points: the best-performing model achieves 80.2\% accuracy and 0.760 macro-F1 with 14{,}040 parameters, while the compact $N_f=25$ configuration uses only 2{,}848 parameters and achieves 75.7\% accuracy, 0.661 macro-F1, and 0.716 MCC on the held-out environment. Analysis of the learned filters and confusion patterns shows that spectral overlap contributes to confusion among water-related events, while the \textit{Door}/\textit{Walker/Crutch} errors also reflect similarities in their transient temporal structure.

1. Introduction

Bathroom acoustic recognition could support assistive monitoring without continuous video, but existing data, interpretability, and model-size constraints complicate deployment. The paper combines a multi-site dataset, an interpretable compact waveform model, and acoustic analysis with design exploration.

1. Introduction

  • Bathroom acoustics offer a less visually intrusive way to monitor activities in private spaces, but suitable public datasets are limited and standard representations can obscure why sounds are confused.The paper also frames resource efficiency as a constraint on practical models.Source: §1 · p. 1
  • SnaanGhar7 contributes 21,387 annotated clips across seven activity-related classes, including a heterogeneous unknown class, with standardized splits and baseline implementations.The corpus is intended for bathroom activities relevant to assistive living.Source: §1 · p. 1
  • SincDPNet pairs a learnable sinc band-pass front end with a depthwise-separable convolutional back end for compact raw-waveform classification.Two frequency parameters per filter make its passbands interpretable in hertz while keeping the filter bank to 2N_f trainable parameters.Source: §1 · pp. 1–2
  • Acoustic overlap analysis relates several dominant class confusions to pairwise similarity, though one class-level difficulty score does not predict the full error ranking.The design study reports all 24 evaluated configurations to expose accuracy–size trade-offs.Source: §1 · p. 2
  • The paper’s contribution is the combination of an interpretable compact waveform model, acoustic analysis, and a dataset enabling study under environment-disjoint evaluation.Bayesian optimization is used as a design tool, not introduced as a new method; SnaanGhar7 is not claimed as the first bathroom acoustic dataset.Source: §1 · p. 2

2. Related Work

Related work frames bathroom recognition around gaps in event coverage, interpretable waveform representations, compact models, efficient design search, and evaluation across rooms. These constraints motivate environment-disjoint testing alongside signal-level analysis.

2. Related Work

  • The related-work categories correspond to constraints of limited data, acoustically similar classes, computational limits, and evaluation outside training domains.These categories span datasets, signal front ends, compact models, optimization, and domain-shift evaluation.Source: §2 · p. 2

2.1 Environmental and Bathroom Acoustic Datasets

  • Bathroom datasets and general environmental corpora often omit assistive-living events such as door and mobility-aid sounds, and explicit heterogeneous non-target classes are uncommon.SnaanGhar7 adds mobility-aid events, a non-target class, and raw-waveform edge baselines.Source: §2.1 · p. 2, Table 1 · p. 2

2.2 Spectral and Raw-Waveform Front Ends

  • Raw-waveform models learn representations directly, while sinc filters constrain each first-layer filter to two cutoff frequencies that can be interpreted in hertz.The cited work motivates structured front ends as a way to reduce free parameters while retaining a direct relation to the signal.Source: §2.2 · p. 3

2.3 Compact Models for Resource-Constrained Audio

  • Compact-audio research uses separable convolutions and other cost-reduction strategies, whereas this study starts with a small parameter budget and jointly considers model size and interpretability.Robustness across recording environments is also identified as important in acoustic monitoring.Source: §2.3 · p. 3

2.4 Multi-Objective Optimization for Edge Models

  • Multi-objective Bayesian optimization is used to search accuracy–complexity and accuracy–latency trade-offs efficiently, not as a methodological contribution.The study uses expected hypervolume improvement to find operating points on a fixed budget.Source: §2.4 · p. 3

2.5 Evaluation Practice and the Gap We Address

  • Evaluation on unseen environments matters because room-specific acoustics can leak across random clip splits; environment-disjoint scores are lower but more realistic for unseen rooms.Signal-level analysis is paired with this protocol to examine whether pre-training class similarities reappear in model confusions.Source: §2.5 · p. 3

3. The SnaanGhar7 Dataset

SnaanGhar7 is a real-world bathroom dataset designed around activity-related sounds relevant to assistive living. Its multi-environment recordings are annotated and partitioned to keep overlapping, room-related windows within a single split.

3. The SnaanGhar7 Dataset

  • SnaanGhar7 contains seven classes: Flush, Shower, Bathroom Tap, Basin Tap, Door, Walker/Crutch, and Unknown Class.Water sounds represent hygiene activities; Door and Walker/Crutch provide entry/exit and mobility-aid cues.Source: §3 · p. 4, Figure 1 · p. 4

3.1 Recording Setup

  • Approximately 55 minutes of recordings were manually annotated per class across five real-world bathroom environments, totaling roughly 385 minutes of original audio.The environments vary in room acoustics, reverberation, plumbing, and background noise.Source: §3 · p. 4, §3.1 · p. 5

3.2 Annotation and Preprocessing

  • Recordings were partitioned by session and environment before overlapping windows were generated, keeping shared audio content and recording-specific characteristics within one partition.Windows contain 30,225 samples at 20 kHz, use a 10,225-sample hop, and overlap by approximately 66%.Source: §3.2 · p. 5

3.3 Dataset Statistics

  • The 21,387 segmented windows have a near-balanced class distribution: Shower is largest at 3,242 and Door smallest at 2,688, for a maximum imbalance ratio of 1.2×.The paper characterizes class imbalance as limited in this benchmark.Source: §3.3 · p. 5

4. Dataset Characterization

The section characterizes the seven classes using model-independent signal descriptors and examines their overlap before training. Sustained water events combine substantial spectral overlap with internal variability, motivating attention to temporal cues and pairwise confusion patterns.

4. Dataset Characterization

  • SnaanGhar7 supports activity-level acoustic classification; a small number of Unknown recordings contain incidental speech treated solely as non-target events.The section’s signal-level analysis uses hand-crafted descriptors before model training.Source: §4 · p. 5

4.1 Signal-Level and Spectral Structure

  • The seven classes differ on RMS, zero-crossing rate, and kurtosis (p<0.001), with impulsive Door and Walker/Crutch clips generally showing higher kurtosis and localized signals.Impulsive classes also occupy higher-centroid and higher-flatness regions, while sustained water classes have lower centroids and smoother spectra.Source: §4.1 · p. 6, Figure 9 · p. 8
  • Water-related classes have large within-class spread, while Flush, Bathroom Tap, and Shower also overlap spectrally and maintain sustained energy profiles.Impulsive Door and Walker/Crutch events instead show sharp onsets and rapid decay, providing a temporal distinction from the water classes.Source: §4.1 · pp. 6, 8, Figure 6 · p. 7, Figure 8 · p. 8

4.2 Inter-Class Overlap and a Difficulty Score

  • Pairwise acoustic overlap better explains specific confusions than the aggregate difficulty score, whose class-level correlation with model error is weak (ρ=−0.07).Flush scores high on overlap and intra-class spread, while impulsive classes are more affected by silence fraction and energy variability; the two-dimensional projection is qualitative.Source: Table 5 · p. 9, §4.2 · p. 9

4.3 Design Implications

  • The analysis is independent of the learned model and proposes that water-class confusion may reflect spectral overlap and internal variability, while onset structure and differing frequency regions may aid discrimination.These are hypotheses about the data’s acoustics, to be examined against learned representations and confusion patterns.Source: §4 · p. 5, §4.1 · p. 8, §4.3 · p. 9

5. SincDPNet Architecture

SincDPNet classifies raw waveforms using a learnable sinc-filter front end and a depthwise-separable convolutional body. The design keeps the frequency bands interpretable while reducing model cost.

5.1 Learnable Sinc Band-Pass Front End

  • Each sinc filter uses two learned frequency parameters to define its passband, so the front end requires only 2N_f trainable parameters independent of kernel length.The learned passbands can be inspected directly in hertz after training.Source: §5.1 · pp. 9, 11, Figure 13 · p. 11

5. SincDPNet Architecture

  • SincDPNet combines a learnable sinc band-pass front end with a depthwise-separable convolutional body, unlike SincNet’s standard-convolution and large fully connected back end.The compact configuration is approximately 28× smaller than the SincNet implementation considered in Table 10.Source: §5 · p. 9, §5.1 · p. 11

5.2 Depthwise-Separable Convolutional Body

  • The front end’s N_f frequency-band channels are reshaped into a time–frequency map, which the depthwise-separable body processes for classification.For k=3 kernels, its parameter cost approaches approximately one ninth of a corresponding standard convolution as C_out increases.Source: §5.2 · p. 12

5.3 Parameter and Complexity Budget

  • In the compact N_f=25 design, most parameters belong to the convolutional body and classification head, rather than the sinc filter bank.The filter-bank parameter count is independent of kernel length L.Source: §5.3 · p. 12, Table 6 · p. 12

6. Multi-Objective Design Optimization: A Practical Case Study

The section uses multi-objective Bayesian optimization as a reproducible design tool to explore SincDPNet accuracy–size and accuracy–latency trade-offs, then selects representative configurations from the resulting Pareto fronts.

6.1 Scope and Framing

  • A 25%-per-class quality-based coreset provides a lower-cost estimate of candidate quality; rank agreement with full-data retraining supports its use for ordering search candidates.Source: §6.1 · p. 13, Figure 14 · p. 13

6.2 Search Space and Objectives

  • The two searches separately maximize validation macro-F1 while minimizing parameter footprint, or maximize validation accuracy while minimizing target-device inference latency.Source: §6.2 · p. 13

6.3 MOBO Formulation

  • qEHVI selects candidates against a dominated reference point using Gaussian-process surrogates, beginning with a Sobol design and updating the Pareto set across evaluations.Source: §6.3 · p. 14

6.4 Acquisition and Kernel Sensitivity

  • The four surrogate-guided settings finish within 1.84 hypervolume units, while the best run exceeds random search by 12.78 units and reaches 90% of its gain after 17 evaluations.Random search does not reach that threshold within the 24-evaluation budget.Source: §6.4 · p. 14, Figure 15 · p. 14, Table 8 · p. 15

6.5 Configuration Selection by Weighted Scoring

  • 0.8942 validation macro-F1 comes from the N_f=59, α=10, five-block model at 54.8 KB; the highest weighted score, 0.8523, comes from an N_f=78 model at 0.8443 macro-F1 and 13.3 KB.Source: Figure 16 · p. 15, §6.5 · p. 15
  • Seven of 24 configurations are Pareto-optimal, spanning 1.4 KB and 0.7622 macro-F1 to 54.8 KB and 0.8942; larger footprints do not always yield higher validation performance.Source: §6.5 · p. 15

7. Experiments and Results

Environment-disjoint evaluation compares SincDPNet with raw-waveform and MFCC baselines, examines model trade-offs, and qualitatively inspects held-out embeddings.

7.1 Experimental Setup

  • Recordings are split by environment, with E01–E03 for training, E04 for validation, and E05 for testing; five-fold leave-one-environment-out evaluation is also reported.Source: §7.1 · pp. 15–16

7.3 Raw-Waveform versus Spectral-Domain Analysis

  • The compact SincDPNet exceeds DPNet while using fewer parameters, and baseline results show no consistent separation between raw-waveform and MFCC inputs.Against DPNet, macro-F1 rises from 0.612 to 0.661 and MCC from 0.645 to 0.716 as parameters fall from 4,944 to 2,848.Source: §7.3 · p. 16, §7.2 · p. 16

7.2 Benchmark Results

  • 80.2% accuracy, 0.847 balanced accuracy, and 0.760 macro-F1: the best-F1 SincDPNet uses 14,040 parameters; the smallest N_f=25 model uses 2,848 parameters and scores 75.7% accuracy, 0.661 macro-F1, and 0.716 MCC.Source: §7.2 · p. 16, Table 10 · p. 17

7.4 Embedding-Space Structure

  • Held-out embeddings show overlapping sustained-water classes and more compact impulsive-class regions across both selected configurations, but the two-dimensional t-SNE view is qualitative and does not establish reliable separation.Source: §7.4 · p. 17, Figure 17 · p. 17

7.5 Ablation Study

The ablations highlight the learned band-pass front end and amplitude augmentation, while filter visualizations relate learned frequency allocation to class responses.

7.5 Ablation Study

  • Removing the band-pass constraint produces the largest front-end ablation reduction; macro-F1 is 0.644 with a fixed sinc bank, 0.615 with matched trainable convolution, and 0.636 with MFCC.The authors caution that seed variation does not establish significance for every pairwise difference.Source: §7.5 · p. 18, Table 11 · p. 18
  • Macro-F1 rises from 0.613 at N_f=16 to 0.672 at N_f=128, with smaller gains beyond 64 filters; removing amplitude augmentation reduces it to 0.625.Longer kernels also show a modest increase: 0.638 at L=101, 0.652 at L=251, and 0.657 at L=401.Source: §7.5 · p. 18
  • The separable body gives 0.651 macro-F1 versus 0.658 for standard convolutions, a difference within seed variation despite substantially more parameters for standard convolutions.Source: §7.5 · p. 18, Table 11 · p. 18

7.6 Learned Filter Interpretation

  • Learned filters concentrate spectral resolution in the low-to-mid range, where water classes carry energy; sustained water classes share active filters, while Basin Tap and impulsive classes show distinct responses.Source: Figure 19 · p. 19, Figure 21 · p. 20, Figure 22 · p. 20

7.7 The Interpretability Bridge

The analysis relates pairwise acoustic overlap to confusion patterns, while showing that temporal structure also matters for transient events and that capacity changes which classes are recovered.

7.7 The Interpretability Bridge

  • Pairwise Bhattacharyya overlap has a modest positive association with off-diagonal confusion (ρ=0.39), whereas class difficulty has little association with compact-model error (ρ=−0.07).Large confusions involve similar water-flow classes and Door/Walker/Crutch; temporal organization remains relevant for transient events.Source: §7.7 · p. 21, Figure 23 · p. 21

7.8 Confusion Structure and Error Characterization

  • Door is classified as Walker/Crutch in 67.7% of cases, while only 1.0% of Walker/Crutch samples are classified as Door; repeated impacts and event boundaries offer a temporal explanation beyond average spectral overlap.Source: §7.8 · p. 21

7.9 Comparison Against the Strongest Baseline

  • SincNet uses approximately 28× more parameters than compact SincDPNet and improves recall for Flush, Door, and Basin Tap, while both retain water-class and Door/Walker/Crutch confusions.Source: §7.9 · p. 21

7.8 Confusion Structure and Error Characterization

  • Basin Tap recall rises from 8.2% in the compact model to 85.2% in the best-F1 model, and Door recall rises from 24.9% to 69.0%.The added capacity chiefly improves classes poorly separated by the compact representation.Source: §7.8 · p. 21, Figure 24 · p. 22

7.9 Comparison Against the Strongest Baseline

  • Across tested architectures, sustained-water embeddings overlap while transient categories form smaller clusters; the two-dimensional visualization does not prove the cause of this pattern.Source: Figure 26 · p. 23, §7.9 · p. 23

8. Deployment-Aware Selection and Hardware Evaluation

The section selects configurations across validation-performance and model-size trade-offs, verifies that conversion preserves predictions, and evaluates single-clip latency on embedded hardware.

8. Deployment-Aware Selection and Hardware Evaluation

  • The weighted-score, accuracy-oriented, and compact Pareto configurations reach validation macro-F1 values of 0.8443, 0.8942, and 0.8016, at 13.3, 54.8, and 11.1 KB, respectively.The compact configuration is retained because no evaluated solution has both a higher macro-F1 and a smaller search-time model size.Source: §8 · p. 23, Table 12 · p. 24

8.2 Deployment on Embedded Hardware

  • All four models achieve a real-time factor well below one on Raspberry Pi Zero 2 W, and the compact model leaves the largest idle margin.The measurements exclude audio capture time.Source: Table 13 · p. 24, §8.2 · p. 24

8.1 Freezing and TFLite Conversion

  • Frozen Keras and FP32 TensorFlow Lite reproduce original Keras predictions exactly on the complete held-out test set; agreement across all three configurations is 100%.The conversion introduces no measurable degradation in these experiments.Source: §8.1 · p. 24

8.2 Deployment on Embedded Hardware

  • Hardware timing compares identical exported models and preprocessing on a Raspberry Pi Zero 2 W and laptop, excluding the fixed 1.5-second audio-capture duration.The objective is to demonstrate lightweight operation on common always-on sensing hardware, not optimize for a specific platform.Source: §8.2 · p. 24

9. Discussion

The discussion connects constrained front ends and compact models to benchmark performance, while relating residual errors to acoustic overlap and temporal structure. It also examines filter interpretability, optimization behavior, and the scope of the resulting design guidance.

Dataset scope

  • SnaanGhar7’s combination of door and mobility-aid sounds with an explicit out-of-distribution class exposes impulsive-event errors that existing bathroom corpora do not capture together.The discussion reports 774 such errors among 1,209 misclassified clips.Source: §9 · p. 25

Error analysis

  • 774 of 1,209 errors are Door↔Walker/Crutch confusions, versus 406 water-flow errors; the two acoustic categories account for 97.6% of errors.No errors were attributed to low signal-to-noise ratio or clipped events; the remaining errors were characterized as real acoustic ambiguity.Source: Table 14 · p. 25, §9 · pp. 25–26

Benchmark comparison

  • SincNet reaches 0.712 macro-F1 and SincDPNet 0.661, compared with DPNet at 0.612 and ACDNet at 0.639; SincDPNet remains Pareto-efficient, not dominant.The SincDPNet–DPNet comparison is at nearly identical sizes: 2,848 versus 4,944 parameters.Source: §9 · p. 25

Architecture and capacity

  • Ablations add 0.014 macro-F1 for band-pass design and 0.027 for learning band-pass edges, while the largest explored architecture reaches only 0.550 macro-F1.Standard convolutions change accuracy by only 0.002 while using nine times more parameters; the discussion cautions that structured front ends do not fully replace capacity.Source: §9 · p. 25

Error analysis

  • The overlap–confusion correlation is ρ=0.39, but Basin Tap ranks mid-difficulty while receiving the worst model performance, so overlap does not fully predict error order.The discussion also notes that which confusion occurs depends on the model’s temporal resolution.Source: §9 · p. 25

Filter interpretability

  • Learned sinc responses distinguish impulsive events and Basin Tap more clearly, while broader low- and mid-frequency responses for sustained water activities align with their confusions.Door and Walker responses are similar; the authors suggest longer temporal context or a dedicated secondary classifier may help separate them.Source: §9 · p. 26

Design exploration

  • Four surrogate-guided optimization variants finish within 0.087% in hypervolume, while random search trails by 0.61%; surrogate variants reach 90% of hypervolume gain within 22 evaluations.The authors frame the value of Bayesian optimization in this small search space as convergence speed and reproducibility, not a dramatically better front.Source: §9 · p. 26

Broader implications

  • The proposed design rule is to analyze class separability before training, reflect it in the front end, and split data by the principal source of deployment variance.That source could be the room, device, site, subject, or recording session.Source: §9 · p. 26

10. Conclusion

The conclusion presents SnaanGhar7 and SincDPNet for waveform-based bathroom activity recognition under an environment-disjoint evaluation protocol. It reports distinct accuracy–size operating points and identifies finer temporal information, broader generalization tests, and constrained-platform deployment as future work.

Conclusion

  • The compact N_f=25 SincDPNet achieves 75.7% accuracy, 0.661 macro-F1, and 0.716 MCC with 2,848 trainable parameters under environment-disjoint evaluation.Recordings are divided by sessions and environments before overlapping segments are formed.Source: §10 · p. 26
  • The highest weighted-score configuration reaches 0.8443 macro-F1 at 13.3 KB, while the highest validation macro-F1 is 0.8942 at 54.8 KB.The conclusion treats these as different Pareto operating points.Source: §10 · p. 26
  • Future work will preserve finer temporal information for Door/Walker confusion, test transfer across rooms and devices, and measure memory, latency, and energy on constrained platforms.The stated deployment targets include microcontroller-class hardware.Source: §10 · p. 27

Ethics Statement

The ethics statement describes participant consent and site permission for the recording collection.

Ethics Statement

Related papers

Loading 2609.34907v1…