Pith. sign in

REVIEW 3 major objections 4 minor 12 references

A CNN-GRU model trained without knowledge distillation achieves 60.25% accuracy on the DCASE 2025 low-complexity acoustic scene classification development set, with 114.2KB memory and 10.9M multiply-accumulate operations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 18:25 UTC pith:SWIXSYO5

load-bearing objection A sensible low-complexity ASC architecture, but the paper never actually reports the proposed model's accuracy, so the central claim is unverifiable. the 3 major comments →

arxiv 2509.09931 v1 pith:SWIXSYO5 submitted 2025-09-12 eess.AS

Acoustic Scene Classification Using CNN-GRU Model Without Knowledge Distillation

Ee-Leng Tan , Jun Wei Yeow , Santi Peksi , Haowen Li , Ziyi Yang , Woon-Seng Gan This is my paper
classification eess.AS
keywords acoustic scene classificationCNN-GRUknowledge distillationlow complexitydata augmentationmel-spectrogramedge inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This technical report presents an acoustic scene classification model for the DCASE 2025 challenge's low-complexity Task 1. The authors claim that a compact CNN-GRU architecture, trained without knowledge distillation, reaches 60.25% average accuracy on the official development dataset while staying within the 128KB and 30MMAC constraints. The model relies on depthwise separable convolutions, channel shuffle, squeeze-and-excite blocks, a frequency-axis GRU, and three augmentation techniques. The paper argues that high performance at low complexity does not necessarily require a teacher-student distillation pipeline.

Core claim

The central claim is that a shallow CNN-GRU model, trained directly on the official development data plus a measured impulse-response dataset for augmentation, can beat the provided baseline and satisfy the challenge's complexity limits. The authors identify the ConvT block (pointwise plus two 1D depthwise convolutions) and a GRU that scans across frequency bins as key design choices. They report 60.25% accuracy on the development dataset with 114.2KB memory and 10.9M MACs, framing the result as evidence that knowledge distillation is not strictly required for data-efficient, low-complexity acoustic scene classification.

What carries the argument

The ConvT block, a module that combines pointwise convolution with two 1D depthwise convolutions applied separately along time and frequency axes, reduces parameters while preserving spectral and temporal structure. A second central mechanism is the GRU configured to process the frequency axis (with frequency bins as features and channels as the sequence), producing features fused with a parallel 1D convolution. Hybrid max/average pooling with squeeze-and-excite blocks and channel shuffle complete the architecture.

Load-bearing premise

The claim rests on the unstated assumption that the reported 60.25% accuracy was computed on a proper held-out evaluation split of the development dataset; the report's tables do not include the proposed model's results, so this assumption is unverified.

What would settle it

Run the proposed model on the official evaluation split of the urban acoustic scene development dataset and compute the macro-average accuracy; if the result differs from 60.25% or the figure only appears when training clips are included, the paper's central empirical claim collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the 60.25% figure holds, low-complexity acoustic scene classification can be accomplished without knowledge distillation, simplifying training pipelines.
  • The model operates within the challenge's 128KB memory and 30MMAC limits, making it deployable on typical edge devices.
  • Using measured impulse responses as the only external data source supports data-efficient training with minimal outside resources.
  • The frequency-axis GRU design offers an alternative to time-sequential modeling that may generalize to other narrowband spectral classification tasks.
  • The combination of SpecAugment, Freq-MixStyle, and impulse-response augmentation improves robustness to unseen recording devices.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's tables only report baseline results; the proposed model's 60.25% accuracy is not shown in a split-by-split breakdown, so the headline number is not directly verifiable from the text alone.
  • If the 60.25% accuracy is reproduced on a standard held-out split, it would suggest that architectural priors and heavy augmentation can substitute for teacher-student distillation in small-model acoustic scene classification.
  • The frequency-axis GRU might transfer to other tasks that classify narrowband spectral signatures, such as machine fault detection or bioacoustic monitoring.
  • A direct comparison against a knowledge-distilled model with identical training data and compute budget would sharpen the claim and clarify the tradeoff.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This technical report describes the SNTL-NTU team's submission to the DCASE 2025 Task 1, an acoustic scene classification system based on a CNN-GRU architecture. The reported model has a memory footprint of 114.2 KB and 10.9M MAC operations, staying within the challenge constraints. The abstract reports an accuracy of 60.25% on the development dataset, comparing favorably to the provided baseline (about 51.9%). The system uses log-mel spectrograms, SpecAugment, Freq-MixStyle, and device-impulse-response augmentation, and is trained without knowledge distillation. The paper includes architecture details, training hyperparameters, and two tables that contain baseline accuracies only.

Significance. If the claimed 60.25% accuracy is reproducible and was obtained on a proper held-out evaluation split, the result is a useful data point for low-complexity acoustic scene classification without knowledge distillation. The reported complexity figures are plausible and the general approach is of interest to the DCASE community. However, the manuscript currently omits the proposed model's own class-wise and device-wise results, and leaves the evaluation split ambiguous. These omissions make the central claim unverifiable from the paper alone. The work would be strengthened by a results table for the proposed model, a precise statement of the evaluation split, and information about the number of independent runs or seeds.

major comments (3)
  1. [Abstract and Section 3 (Tables I and II)] The central accuracy claim of 60.25% appears only in the abstract. Tables I and II are captioned 'Baseline' and report only baseline class-wise and device-wise accuracies. No table or figure reports the proposed model's class-wise or device-wise performance. Without this table, the reader cannot verify the claimed accuracy, compare the model against the baseline, or check per-class behavior. This is load-bearing because the entire contribution rests on the 60.25% figure.
  2. [Abstract and Section 3] The phrase 'using the development dataset' is ambiguous. In DCASE Task 1, the development set is split into training and evaluation partitions, and an accuracy computed on the full development set (including training clips) would be invalid. The manuscript does not state whether the 60.25% figure comes from the official evaluation split, a held-out subset, or the full development set. Please clarify the exact evaluation protocol; if the reported number is from the full development set, it would invalidate the central claim.
  3. [Section 3 (training details)] The paper reports a single accuracy value without error bars, confidence intervals, or the number of training runs. Given that the baseline is reported as a single number too, the claimed improvement of roughly 8.4 percentage points should be accompanied by at least a statement on variance or replicate runs. Without this, the reader cannot distinguish a robust result from a favorable seed or data-split artifact.
minor comments (4)
  1. [Section 2.1 and Section 3] The preprocessing description says a window length of 0.18 s and '17% overlap', while Section 3 gives window length 8192, hop length 1364, and FFT size 8192. At a 44.1 kHz sampling rate, 1364/8192 corresponds to about 16.7% hop and therefore about 83.3% overlap, not 17% overlap. Please correct the text to avoid a contradiction.
  2. [Section 2.1 and References] The citation to librosa is given as [3], but reference [3] in the bibliography is the MicIRP link and librosa is reference [4]. Many other reference numbers also appear misaligned with the list. Please renumber citations carefully.
  3. [Abstract] There is a typo: 'muliply-and-accumulate' should be 'multiply-and-accumulate'.
  4. [Section 2.3 and Figure 2] The GRU hidden size, the exact channel counts in the ConvT blocks, and the number of GRU layers are not specified in the text. The figure shows channel counts, but the GRU's hidden dimension and whether it is unidirectional/bidirectional are not stated. Adding these details would improve reproducibility.

Circularity Check

0 steps flagged

No circularity: empirical system description with measured accuracy; reporting omissions are not circular reasoning.

full rationale

This paper is a DCASE challenge system description, not a derivation of a predicted quantity from assumptions. The central claims are architectural (114.2 KB memory, 10.9M MACs) and empirical (60.25% accuracy on the development dataset), both reported as measured outcomes rather than quantities defined in terms of fitted parameters. There is no step where an input assumption already contains the result: the accuracy is not constructed from the model definition, and the memory/MAC counts are direct properties of the architecture. The paper contains no self-citations by the present authors that are load-bearing, no imported uniqueness theorem, and no ansatz smuggled in via citation. The most notable issue is that Tables I and II only report baseline accuracies and not the proposed model's class-wise or device-wise results, making the 60.25% figure hard to verify from the manuscript. That is a reporting/reproducibility concern, not circularity. The evaluation split ambiguity ('using the development dataset') is likewise an accountability issue, not a circular dependency. No passage in the manuscript asserts a limitation or admits a circular step. The analysis therefore finds no significant circularity and assigns a score of 0.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

This is an empirical engineering report. The central claim (accuracy and complexity) depends on a hand-designed architecture, manually chosen hyperparameters, and public data. No new entities are postulated. The main unstated assumptions are the evaluation protocol and the sufficiency of the mel-spectrogram configuration.

free parameters (4)
  • Mel-spectrogram window/hop/FFT/mel-bin settings = 8192/1364/8192/256
    Chosen by hand to balance spectral resolution and computational cost; the 10.9M MAC count depends on the 256x33 input shape.
  • ConvT channel widths = 7, 24, 72, 72, 144, 28, 10
    Designed manually; no ablation is reported.
  • Dropout rate = 30%
    Set to 0.3, a standard heuristic; effect on accuracy not analyzed.
  • Training epochs = 150
    Selected for the challenge budget; not shown to be optimal.
axioms (3)
  • domain assumption TAU Urban Acoustic Scene 2022 Mobile development dataset labels are correct and the standard splits are used.
    The evaluation rests on the dataset; the paper does not specify which split produced 60.25%.
  • domain assumption Mel-spectrogram features with 256 bins and STFT window 8192 are sufficient for classification.
    Chosen based on average spectra in Fig. 1, not validated by an ablation.
  • domain assumption The DCASE complexity constraints (128 KB memory, 30 MMAC) are the appropriate evaluation target.
    The paper optimizes for these constraints; no comparison at other complexity thresholds.

pith-pipeline@v1.3.0-alltime-deepseek · 3554 in / 8183 out tokens · 78883 ms · 2026-08-04T18:25:43.279368+00:00 · methodology

0 comments
read the original abstract

In this technical report, we present the SNTL-NTU team's Task 1 submission for the Low-Complexity Acoustic Scenes and Events (DCASE) 2025 challenge. This submission departs from the typical application of knowledge distillation from a teacher to a student model, aiming to achieve high performance with limited complexity. The proposed model is based on a CNN-GRU model and is trained solely using the TAU Urban Acoustic Scene 2022 Mobile development dataset, without utilizing any external datasets, except for MicIRP, which is used for device impulse response (DIR) augmentation. The proposed model has a memory usage of 114.2KB and requires 10.9M muliply-and-accumulate (MAC) operations. Using the development dataset, the proposed model achieved an accuracy of 60.25%.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references

  1. [1]

    INTRODUCTION In Task 1 of the DCASE Challenge 202 5, acoustic scene classifi- cation (ASC) is employed to classify 10 acoustic scenes from 12 cities based on 1 -second audio samples. To align ASC with the performance of typical edge devices, Task 1 [1] of the DCASE Challenge 2023 has imposed the following system complexity constraints: • Maximum memory al...

    2023

  2. [2]

    Preprocessing The TAU urban acoustic scene 2022 mobile dataset contains recordings of 10 acoustic scenes in 12 European cities

    PROPOSED SYSTEM 2.1. Preprocessing The TAU urban acoustic scene 2022 mobile dataset contains recordings of 10 acoustic scenes in 12 European cities. These recordings are captured using four devices and synthetic data for 11 devices was generated usi ng the recordings. Each 1 sec audio sample is captured with a sampling frequency of 44.1 kHz sam- pling rat...

    2022

  3. [3]

    The window length, hop length, FFT size , and number of mel bins are 8192, 1364, 8192, and 256, respectively

    RESULTS AND SUBMISSION For all splits, t he proposed model w as trained for 150 epochs with a batch size of 256 using the ADAM optimizer with a learning rate adjusted by the cosine schedule with ramp -up. The window length, hop length, FFT size , and number of mel bins are 8192, 1364, 8192, and 256, respectively. The results of the provided baseline and s...

  4. [4]

    The proposed model is based on CNN and is trained solely on the TAU Urban Acoustic Scene 2022 Mobile development dataset

    CONCLUSIONS In this technical report, w e described the SNTL-NTU submis- sions to task 1 of the DCASE 202 5 challenge. The proposed model is based on CNN and is trained solely on the TAU Urban Acoustic Scene 2022 Mobile development dataset. Detection and Classification of Acoustic Scenes and Events 2025 Challenge Table I Class-Wise Accuracies of Baseline ...

    2022

  5. [5]

    ACKNOWLEDGEMENT This research is supported by the Ministry of Education, Singa- pore, under its Academic Research Fund Tier 2 (MOE - T2EP20221-0014)

  6. [6]

    Data-efficient l ow-complexity acoustic sce- ne classification in the DCASE 2024 challenge,

    Florian Schmid, Paul Primus, Toni Heittola, Annamaria Mesaros, Irene Martín -Morató, Khaled Koutini, and Ger- hard Widmer, “Data-efficient l ow-complexity acoustic sce- ne classification in the DCASE 2024 challenge,” 2024

    2024

  7. [7]

    Heittola, A

    T. Heittola, A. Mesaros, and T. Virtanen, “Acoustic scene classification in DCASE 2020 challenge: generalization across devices and low complexity solutions. In Proceedings of the Detection and Classification of Acoustic Scenes and Events 2020 Workshop (DCASE2020),” 56–60. 2020

    2020

  8. [8]

    https://micirp.blogspot.com/?m=1

  9. [9]

    librosa: Audio and music signal analysis in python,

    B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and music signal analysis in python,” SciPy, 2015

    2015

  10. [10]

    SpecAugment: A simple data augmenta- tion method for automatic speech recognition,

    D. S. Pa rk, et al., “SpecAugment: A simple data augmenta- tion method for automatic speech recognition,” Interspeech, pp. 2613-2617, 2019

    2019

  11. [11]

    Distilling the knowledge of transformers and CNNs with CP-mobile. In Proceedings of the Detection and Classification of Acoustic Scenes and Events 2023 Work- shop (DCASE2023),

    F. Schmid, T. Morocutti, S. Masoudian, K. Koutini, and G. Widmer, “ Distilling the knowledge of transformers and CNNs with CP-mobile. In Proceedings of the Detection and Classification of Acoustic Scenes and Events 2023 Work- shop (DCASE2023),” 161–165. 2023

    2023

  12. [12]

    De- vice robust acoustic scene classification via impulse re- sponse augmentation,

    T. Morocutti, F. Schmid, K. Koutini , and G. Widmer, “De- vice robust acoustic scene classification via impulse re- sponse augmentation,” in 31st EUSIPCO, 2023

    2023