Infant Cry Emotion Recognition Using Improved ECAPA-TDNN with Multiscale Feature Fusion and Attention Enhancement

Haolin Yu; Junyu Zhou; Yanxiong Li

REVIEW 4 major objections 5 minor 26 references

Infant Cry Emotion Recognition Using Improved ECAPA-TDNN with Multiscale Feature Fusion and Attention Enhancement

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read An ECAPA-TDNN variant with multiscale feature fusion and attention enhancement claims 82.20 percent accuracy on six-class infant cry emotion recognition.

desk verdict A plausible incremental architecture paper whose headline accuracy rests on a single split that may leak speaker identity; worth reviewing, but the evaluation protocol needs to be fixed before the number is taken seriously. read the letter →

arxiv 2506.18402 v1 pith:P3425IQQ submitted 2025-06-23 eess.AS

Junyu Zhou , Yanxiong Li , Haolin Yu This is my paper · ORCID

classification eess.AS

keywords infantcryemotionrecognitionECAPA-TDNNmulti-scalefeaturefusionattentionmechanismchanneldifferentialclassificationlightweightmodels

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a modified ECAPA-TDNN—a time-delay neural network with emphasized channel attention—recognizes six types of infant cry emotion more accurately than the original network. The modifications add multiscale feature fusion, a temporal-channel attention interaction, and differential channel attention. On a public dataset of 3000 short cry clips, the improved model reports 82.20 percent accuracy with 1.43 MB of parameters and 0.32 Giga FLOPs per sample, beating the ECAPA-TDNN baseline at 73.38 percent and the ResNet18 baseline at 60.07 percent. The result matters because automatic cry-emotion recognition could support parenting and medical monitoring, and the paper's additions target the subtle time-frequency structure that makes cries hard to classify. The comparison is made on a single 8:2 train-test split.

What carries the argument

The load-bearing object is the improved ECAPA-TDNN architecture, specifically three inserted modules. The Multi-scale Channel Attention (MCA) module runs parallel dilated one-dimensional convolutions with kernel sizes 3, 5, and 7 plus a max-pooled branch, concatenates them along the channel dimension, and recalibrates the result with a squeeze-and-excitation gate; this widens the receptive field without large extra computation. The Residual Squeeze-and-Excitation (RSE) module adds a residual connection around a squeeze-and-excitation block and fuses temporal attention with channel attention through their outer product, coupling time and frequency cues. The differential attention module computes two softmax attention maps from split query and key projections and subtracts them with a learnable scalar, canceling shared attention noise and amplifying emotion-relevant channels. The architecture also replaces attentive statistics pooling with global average pooling and the AAM-Softmax loss with plain Softmax. These modules are what capture localized events such as pain spikes and longer patterns such as hunger rhythms, and what integrate temporal and spectral information.

What would settle it

Retrain the proposed and baseline models on the same dataset under a speaker-disjoint split, for example leave-one-infant-out, and repeat with several random 8:2 splits. If accuracy on unseen infants falls to near the 73.38 percent ECAPA-TDNN baseline, the reported 82.20 percent advantage is mostly split-specific identity leakage rather than emotion recognition.

Watch

Extended reading notes

Core claim

The paper's central claim is that the ECAPA-TDNN's single-scale convolutions, weak interlayer fusion, and isolated temporal and channel attention can be addressed by three cooperative modules, and that the resulting network outperforms both the original ECAPA-TDNN and ResNet18 on six-class infant cry emotion recognition. The improved network reaches 82.20 percent accuracy with 1.43 MB parameters and 0.32 Giga FLOPs on the public 3000-clip dataset labeled awake, diaper, hug, hungry, sleepy, and uncomfortable. Ablations attribute the largest accuracy contribution to the multiscale channel attention module: removing it drops accuracy from 82.20 to 78.37 percent, while removing the residual squeeze-excitation module gives 78.54 percent and removing differential attention gives 78.70 percent. The paper concludes that the three modules are complementary and that the accuracy advantage over the baselines comes from better integration of multiscale and time-frequency information.

Load-bearing premise

The result rests on a single 8:2 train-test split that the paper does not show to be speaker-disjoint, so if clips from the same infant appear in both parts the model could be learning who is crying rather than what the cry means.

Editorial extensions

If this is right

On the six-class dataset used here, the proposed model raises accuracy from 73.38 percent (ECAPA-TDNN) and 60.07 percent (ResNet18) to 82.20 percent.
The model stays small at 1.43 MB parameters and 0.32 Giga FLOPs per sample, which is heavier than the original ECAPA-TDNN but far lighter than ResNet18, so edge-device deployment remains a plausible goal.
Ablation results indicate that the multiscale channel attention module contributes the most to the gain, and that all three modules complement one another.
The remaining hardest decision boundary in the proposed model is between the awake and hungry classes, with 12.15 percent confusion.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: because the paper reports one fixed 8:2 split without showing that clips from the same infant cannot appear in both training and test sets, the 82.20 percent figure is best read as an upper bound; a speaker-disjoint split or repeated cross-validation would reveal how much of the gain is emotion recognition rather than identity matching.
Editorial inference: the multiscale channel attention and differential attention modules are not cry-specific, so the same recipe could be tested on other short-audio paralinguistic tasks such as adult emotion or stress recognition, where time-frequency structure and scarce data pose similar challenges.
Editorial inference: since the paper replaces the pooling layer and the loss function at the same time as adding the three modules, a clean test would keep one change fixed and vary the others, isolating each contribution more sharply than the reported ablations do.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is an incremental architecture paper that does what it says, but the headline accuracy is only as good as a single unpartitioned split, and the paper never shows the split is leak-proof.

The actual contribution is a specific combination of three existing attention/fusion mechanisms—multi-scale dilated convolution with squeeze-excitation, residual SE, and differential attention—grafted onto ECAPA-TDNN. That combination hasn't been benchmarked on this public infant-cry dataset before, and the authors report parameter count and FLOPs alongside accuracy, release code, and run a clean ablation showing each module contributes something. Credit where it's due: the ablation is simple but informative, and the conclusion honestly notes the model is heavier than the base ECAPA-TDNN and that spatial dependencies are under-exploited. That is the right way to report incremental work.

The soft spot is the evaluation protocol, and it is the whole ballgame. Section 3.1 says only that the dataset has 3000 clips and the train/test ratio is 8:2. It never says whether the split is clip-level random, stratified by class, or speaker-disjoint. With infant cries, if clips from the same infant or the same recording session appear in both partitions, the model can learn speaker or recording identity instead of cry emotion. The reported 82.20% could be inflated by exactly that kind of leakage. Even without leakage, 600 test clips make the ablation gaps (82.20 vs 78.37–78.70) roughly 20–25 samples; that is within sampling noise unless the split is fixed and the result is stable across repeats. The paper also doesn't compare against previously published numbers on the same dataset, only against its own reimplementations of ECAPA-TDNN and ResNet18. Those baselines are fine as internal controls, but they don't tell us where the proposed model sits relative to the literature.

None of this means the work is wrong. The architecture is reasonable, the components are all cited from prior work, and the code is promised. It means the central accuracy claim is not yet established. I'd send this to peer review, but the reviewers should demand a speaker-disjoint split (or at minimum a clear statement of the split unit), repeated runs with error bars or confidence intervals, and a direct comparison to any published results on the iFLYTEK/USTC dataset. If those come back okay, this is a solid incremental data point for the infant-cry recognition community.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes an improved ECAPA-TDNN for six-class infant cry emotion recognition, adding a multi-scale channel attention (MCA) module, a residual squeeze-and-excitation (RSE) module, and a differential-attention channel recalibration stage to the standard ECAPA-TDNN. On an iFLYTEK/USTC public dataset of 3,000 clips, the method is reported to reach 82.20% accuracy with 1.43 MB parameters and 0.32 Giga FLOPs, outperforming the ECAPA-TDNN (73.38%) and ResNet18 (60.07%) baselines implemented by the authors. Ablation experiments in Table 1 attribute the gain to each of the three modules, and the conclusion acknowledges the model's higher complexity than ECAPA-TDNN and the lack of spatial-spectral attention.

Significance. The contribution is an incremental but useful empirical architecture study. The ablation design is sensible, the comparison includes a parameter/FLOPs trade-off, and the code is released at a public repository, which supports reproducibility. If the reported accuracy genuinely reflects generalization to unseen infants, the method offers a competitive accuracy-complexity point for infant cry emotion recognition. However, the significance is tempered because the paper evaluates on a single split with no repeated runs, no speaker-disjoint guarantee, and no statistical significance testing, so the magnitude of the improvement over baselines is uncertain. The manuscript honestly states its own limitations, including higher complexity than ECAPA-TDNN and underutilized spatial dependencies.

major comments (4)

[Section 3.1] The evaluation protocol is underspecified in a load-bearing way. The paper states only that the dataset contains 3,000 clips and that the train/test ratio is 8:2; it does not state whether the split is speaker-disjoint, stratified by class, or performed at the clip level. Because the central claim is generalization to unseen infant cry recordings, an evaluation in which clips from the same infant appear in both partitions could inflate the 82.20% accuracy by letting the model exploit speaker or recording identity. Please specify the split unit, number of subjects, and class distribution, and either provide a speaker-disjoint split or justify why a clip-level random split is sufficient for this dataset.
[Table 1 and Section 3.3] The ablation conclusions are reported without uncertainty quantification. With 600 test clips, the differences in Table 1 (e.g., 82.20% vs. 78.37%, a difference of about 23 clips) are within the range expected from sampling noise, so the claim that each module contributes meaningfully needs repeated runs, confidence intervals, or a statistical significance test. Similarly, the headline gain of 82.20% over the ECAPA-TDNN baseline (73.38%) should be accompanied by the variability of these estimates before it is presented as a definitive advantage.
[Section 3.3, Table 2] The comparison is limited to the two baselines implemented by the authors. The paper does not compare with published deep-learning methods on the same iFLYTEK/USTC dataset or with standard speech emotion recognition embeddings, and it does not report whether the baselines use the same preprocessing, augmentation, and optimization budget. Without this context, the headline accuracy advantage over ECAPA-TDNN and ResNet18 cannot be assessed as a claim about state of the art.
[Section 2.2, Eqs. (2)-(5)] There is a dimension inconsistency in the specification of the MCA module. The input X is introduced as X in R^(C×T), but Eq. (3) writes M in R^(H×W×C) after a 1×1 convolution and MaxPool, which is incompatible with the concatenation in Eq. (2) into R^(4C×T). In addition, Eq. (5) defines W1 in R^((3C/r)×3C) and W2 in R^(3C×(3C/r)) while z in Eq. (4) is R^(4C); the matrix dimensions do not match. If these are typographical errors, they should be corrected because the MCA module is central to the proposed method.

minor comments (5)

[Figures] Figure 4 is missing from the text; Figures 5-7 follow Figure 3, so the figure numbers should be renumbered or a reference to Figure 4 should be added.
[Section 3.2] The learning rate appears as '2ିହ' in the text; it should be typeset as 2^-5.
[Section 3.2] The preprocessing description says '13 filters are adopted for spectral feature extraction' but does not define the filterbank, window length, hop size, or whether the input is a raw waveform or a log-mel spectrogram; these details are needed for reproducibility.
[Section 3.3] The sentence 'That is, it doesn't and does have advantage over the ECAPA-TDNN and the Resnet18 in complexity, respectively' is confusing; it should be rewritten as 'the proposed model has higher complexity than ECAPA-TDNN but lower complexity than ResNet18.'
[Section 1] The abstract and introduction present the 82.20% accuracy without qualification; it would be more precise to say 'on the test split used in this study' given the single-split evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the improved ECAPA-TDNN is an empirical architecture comparison with all components and baselines drawn from external prior work.

full rationale

The paper's central claim is an empirical accuracy result (82.20%) on a public dataset. The proposed modules (MCA, RSE, differential attention) are defined by explicit equations that transform feature maps and do not reference the target labels or the reported accuracy; there is no derivation of the accuracy from the model equations. The evaluation is a standard train/test procedure with an 8:2 split, and no parameter or module is fitted to the test labels. The baselines (ECAPA-TDNN, ResNet18) are re-implemented under the same protocol, so the accuracy gain is a measured comparison, not a reduction. All citations point to external prior work ([17]-[21]); there is no self-citation chain, no imported uniqueness theorem, and no fitted input renamed as prediction. The stated limitations (higher complexity than ECAPA-TDNN, underutilized spatial dependencies) show the authors do not claim a forced outcome. The only substantive concern—whether the 8:2 split is speaker-disjoint—is a validity or potential data-leakage issue, not a circularity step, and does not affect the circularity score.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The model modules are computational components built from published attention and multi-scale convolution ideas. The empirical claim rests on the dataset labels, the split protocol, and hand-chosen hyperparameters listed above.

free parameters (6)

Learning rate = 2^-5
Set to 2^-5 without a sensitivity study; directly affects optimization and final accuracy.
Batch size = 64
Chosen for training on an RTX 3060; no sensitivity analysis reported.
Training epochs = 700
Stopping criterion not described; final accuracy could depend on this choice.
Number of spectral filters = 13
The text says 13 filters are adopted for spectral feature extraction, but the exact feature type is not specified.
Number of channels = 128
Network width used for all experiments; effect on accuracy not analyzed.
Attention reduction ratio r = not reported
Used in SE and MCA gating (Equations 5 and 8); value is not stated in the paper.

assumptions (3)

domain assumption The emotion labels in the iFLYTEK/USTC dataset are correct ground truth.
Section 3.1 introduces the six emotion classes with no information about labeling procedure, inter-rater agreement, or expert validation; all accuracy numbers depend on these labels.
domain assumption The 8:2 train-test split prevents speaker or recording identity leakage.
Section 3.1 gives only the ratio, not whether clips from the same infant are kept in one partition; if not, the model can memorize voice identity instead of emotion.
domain assumption Spectral features with 13 filters capture emotion-relevant information.
Section 3.2 states 13 filters are used for spectral feature extraction without specifying the feature type or justifying the filter count.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Infant Cry Emotion Recognition Using Improved ECAPA-TDNN with Multiscale Feature Fusion and Attention Enhancement." pith.science (2026). https://pith.science/paper/P3425IQQ

@misc{pith2026250618402,
  author       = {Pith},
  title        = {Pith review of: Infant Cry Emotion Recognition Using Improved ECAPA-TDNN with Multiscale Feature Fusion and Attention Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3425IQQ}},
  note         = {Machine review of arXiv:2506.18402}
}

read the original abstract

Infant cry emotion recognition is crucial for parenting and medical applications. It faces many challenges, such as subtle emotional variations, noise interference, and limited data. The existing methods lack the ability to effectively integrate multi-scale features and temporal-frequency relationships. In this study, we propose a method for infant cry emotion recognition using an improved Emphasized Channel Attention, Propagation and Aggregation in Time Delay Neural Network (ECAPA-TDNN) with both multi-scale feature fusion and attention enhancement. Experiments on a public dataset show that the proposed method achieves accuracy of 82.20%, number of parameters of 1.43 MB and FLOPs of 0.32 Giga. Moreover, our method has advantage over the baseline methods in terms of accuracy. The code is at https://github.com/kkpretend/IETMA.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

[1]

Dunstan’s theory identifies primal, reflexive vocalizations in infants aged 0-3 months linked to specific physiological needs, verified in later studies [1]

Introduction Infant cries encode biologically significant information through acoustic features like frequency, duration, and timbre, correlating with emotional/physical states and influenced by individual variability and environmental noise. Dunstan’s theory identifies primal, reflexive vocalizations in infants aged 0-3 months linked to specific physiolo...

work page
[2]

Figure 1 illustrates the frameworks of both ECAPA- TDNN and improved ECAPA-TDNN

Method 2.1 Method Framework The framework of the improved ECAPA-TDNN is the same as that of the ECAPA-TDNN, except for three synergistic modules. Figure 1 illustrates the frameworks of both ECAPA- TDNN and improved ECAPA-TDNN. In the improved ECAPA-TDNN, the modules of Residual Squeeze-and-Excitation (RSE), Multi-scale Channel Attention (MCA) and differen...

work page
[3]

awake”, “diaper

Experiments 3.1 Experimental Datasets The experiments are conducted on a public dataset which is released by iFLYTEK and the University of Science and Technology of China in 2020. This dataset contains a total of 3000 audio clips. The duration of each audio clip ranges from 3 to 5 seconds. There are six types of infant cry emotions in this dataset, namely...

work page 2020
[4]

It integrates multi- scale feature fusion module and bidirectional temporal-channel co-attention modules

Conclusions In this paper, we design a model of the improved ECAPA- TDNN for infant cry emotion recognition. It integrates multi- scale feature fusion module and bidirectional temporal-channel co-attention modules. Experimental results demonstrate that the proposed model achieves the accuracy score of 82.20% which is higher than that obtained by two state...

work page
[5]

Acknowledgement This work was supported by the national natural science foundation of China (62371195, 62111530145, 61771200), the exchange project of the 10th Meeting of the China-Croatia Science and Technology Cooperation Committee (10-34), the provincial undergraduate training program for innovation and entrepreneurship (S202410561216), and Guangdong p...

work page
[6]

Dunstan baby language classification with CNN,

C. A. Bratan, M. Gheorghe, I. Ispas, E. Franti, M. Dascalu, S. M. Stoicescu, I. Roșca, F. Gherghiceanu, D. Dumitrache, and L. Nastase, "Dunstan baby language classification with CNN," in Proc. of International Conference on Speech Technology and Human-Computer Dialogue (SpeD), Bucharest, Romania, 2021, pp. 167-171

work page 2021
[7]

Infant crying detection in real-world environments,

X. Yao, M. Micheletti, M. Johnson, E. Thomaz, and K. de Barbaro, "Infant crying detection in real-world environments," in Proc. of IEEE International Conference on Acoustics, Speech and Signal Processing, Singapore, 2022, pp. 131-135

work page 2022
[8]

Is your baby fine at home? Baby cry sound detection in domestic environments,

T. Khandelwal, R. K. Das, and E. S. Chng, "Is your baby fine at home? Baby cry sound detection in domestic environments," in Proc. of Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Chiang Mai, Thailand, 2022, pp. 275-280

work page 2022

Show all 26 references

[9]

Real-time baby crying detection in the noisy everyday environment,

L. S. Foo, W.-S. Yap, Y. C. Hum, Z. Kadim, H. W. Hon, and Y. K. Tee, "Real-time baby crying detection in the noisy everyday environment," in Proc. of IEEE Control and System Graduate Research Colloquium (ICSGRC), Shah Alam, Malaysia, 2020, pp. 26-31

2020
[10]

Spectrogram and LSTM based infant cry detection method for infant wellness monitoring systems,

S. P. Narayanan, M. S. Manikandan, and L. R. Cenkeramaddi, "Spectrogram and LSTM based infant cry detection method for infant wellness monitoring systems," in Proc. of International Conference on Human System Interaction (HSI), Paris, France, 2024, pp. 1-4

2024
[11]

A machine learning approach to classify biomedical acoustic features for baby cries,

G. Aggarwal, K. Jhajharia, J. Izhar, M. Kumar, and L. Abualigah, "A machine learning approach to classify biomedical acoustic features for baby cries," Journal of Voice, vol. 38, no. 2, pp. 245- 253, 2023

2023
[12]

Infant cry language analysis and recognition: an experimental approach,

L. Liu, W. Li, X. Wu, and B. X. Zhou, "Infant cry language analysis and recognition: an experimental approach," IEEE/CAA Journal of Automatica Sinica, vol. 6, no. 3, pp. 778-788, May 2019

2019
[13]

A multistage heterogeneous stacking ensemble model for augmented infant cry classification,

V.R. Joshi, K. Srinivasan, P.M.D.R. Vincent, V. Rajinikanth, and C.Y. Chang, "A multistage heterogeneous stacking ensemble model for augmented infant cry classification," Frontiers in Public Health, vol. 10, p. 819865, Mar. 2022

2022
[14]

Infant cry classification using transfer learning,

G. Anjali, S. Sanjeev, A. Mounika, G. Suhas, G. P. Reddy, and Y. Kshiraja, "Infant cry classification using transfer learning," in Proc. of IEEE Region 10 Conference on TENCON, Hong Kong, 2022, pp. 1-7,

2022
[15]

Infant speech emotion recognition based on channel attention mechanism with ResNet- BiLSTM,

X. Shen, B. Lv, T. Liu, and Q. Cheng, "Infant speech emotion recognition based on channel attention mechanism with ResNet- BiLSTM," in Proc. of International Conference on Information Science, Parallel and Distributed Systems (ISPDS), Guangzhou, China, 2024, pp. 54-57

2024
[16]

An automated mood analysis of crying infants through sound recognition using hybrid deep learning,

S. G. A, G. S, G. Tharagarani, S.P, and S. B, "An automated mood analysis of crying infants through sound recognition using hybrid deep learning," in Proc. of International Conference on Smart Technologies and Systems for Next Generation Computing (ICSTSN), Villupuram, India, ...

2024
[17]

Emotion recognition of infant crying sounds using convolutional recurrent neural network with multi-scale joint attention mechanism,

Y. Liu, B. Lv, S. Xu, and X. Shen, "Emotion recognition of infant crying sounds using convolutional recurrent neural network with multi-scale joint attention mechanism," in Proc. of International Conference on Information Systems Engineering (ICISE), Dalian, China, 2023, pp. 615-619

2023
[18]

A multi-scale convolutional attention neural network based on residual block downsampling for infant cry classification and detection,

J. Yang, Z. Zhang, J. Li, and C. Lin, "A multi-scale convolutional attention neural network based on residual block downsampling for infant cry classification and detection," in Proc. of International Conference on Internet of Things, Automation and Artificial Intelligence (Io...

2024
[19]

A review of infant cry analysis and classification,

C. Ji, T. B. Mudiyanselage, Y. Gao, and Y. Pan, "A review of infant cry analysis and classification," EURASIP Journal on Audio, Speech, and Music Processing, vol. 2021, no. 1, Art. no. 8, 2021

2021
[20]

Defining and distinguishing infant behavioral states using acoustic cry analysis: Is colic painful ?,

J. J. Parga, S. K. Lewis, and M. H. Goldstein, "Defining and distinguishing infant behavioral states using acoustic cry analysis: Is colic painful ?," Pediatric Research, vol. 87, no. 3, pp. 440-447, 2020

2020
[21]

Self-supervised learning for infant cry analysis,

A. Gorin, C. Subakan, S. Abdoli, J. Wang, S. Latremouille, and C. C. Onu, "Self-supervised learning for infant cry analysis," in Proc. of ICASSP Workshop on Safety and Security in Speech and Biomedical Signal Processing (SASB), IEEE, 2023, pp. 1-5

2023
[22]

ECAPA- TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,

B. Desplanques, J. Thienpondt, and K. Demuynck, "ECAPA- TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification," in Proc. of INTERSPEECH, 2020, pp. 3830-3834

2020
[23]

CryCeleb: A speaker verification dataset based on infant cry sounds,

D. Budaghyan, C. C. Onu, A. Gorin, C. Subakan, and D. Precup, "CryCeleb: A speaker verification dataset based on infant cry sounds," in Proc. of IEEE International Conference on Acoustics, Speech and Signal Processing, Seoul, Korea, 2024, pp. 11966-11970

2024
[24]

Improving speaker representations using contrastive losses on multi-scale features,

S. Dixit, A. Jain, and R. Singh, "Improving speaker representations using contrastive losses on multi-scale features," arXiv preprint arXiv:2410.05037, 2024

2024 arXiv
[25]

Differential transformer,

T. Ye, L. Wang, and H. Li, "Differential transformer," arXiv preprint arXiv:2410.05258, 2024

2024 arXiv
[26]

Classification of Infant Crying Sounds Using SE-ResNet-Transformer,

F. Li, C. Cui, and Y. Hu, "Classification of Infant Crying Sounds Using SE-ResNet-Transformer," Sensors, vol. 24, no. 20, pp. 6575, 2024

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Dunstan’s theory identifies primal, reflexive vocalizations in infants aged 0-3 months linked to specific physiological needs, verified in later studies [1]

Introduction Infant cries encode biologically significant information through acoustic features like frequency, duration, and timbre, correlating with emotional/physical states and influenced by individual variability and environmental noise. Dunstan’s theory identifies primal, reflexive vocalizations in infants aged 0-3 months linked to specific physiolo...

work page

[2] [2]

Figure 1 illustrates the frameworks of both ECAPA- TDNN and improved ECAPA-TDNN

Method 2.1 Method Framework The framework of the improved ECAPA-TDNN is the same as that of the ECAPA-TDNN, except for three synergistic modules. Figure 1 illustrates the frameworks of both ECAPA- TDNN and improved ECAPA-TDNN. In the improved ECAPA-TDNN, the modules of Residual Squeeze-and-Excitation (RSE), Multi-scale Channel Attention (MCA) and differen...

work page

[3] [3]

awake”, “diaper

Experiments 3.1 Experimental Datasets The experiments are conducted on a public dataset which is released by iFLYTEK and the University of Science and Technology of China in 2020. This dataset contains a total of 3000 audio clips. The duration of each audio clip ranges from 3 to 5 seconds. There are six types of infant cry emotions in this dataset, namely...

work page 2020

[4] [4]

It integrates multi- scale feature fusion module and bidirectional temporal-channel co-attention modules

Conclusions In this paper, we design a model of the improved ECAPA- TDNN for infant cry emotion recognition. It integrates multi- scale feature fusion module and bidirectional temporal-channel co-attention modules. Experimental results demonstrate that the proposed model achieves the accuracy score of 82.20% which is higher than that obtained by two state...

work page

[5] [5]

Acknowledgement This work was supported by the national natural science foundation of China (62371195, 62111530145, 61771200), the exchange project of the 10th Meeting of the China-Croatia Science and Technology Cooperation Committee (10-34), the provincial undergraduate training program for innovation and entrepreneurship (S202410561216), and Guangdong p...

work page

[6] [6]

Dunstan baby language classification with CNN,

C. A. Bratan, M. Gheorghe, I. Ispas, E. Franti, M. Dascalu, S. M. Stoicescu, I. Roșca, F. Gherghiceanu, D. Dumitrache, and L. Nastase, "Dunstan baby language classification with CNN," in Proc. of International Conference on Speech Technology and Human-Computer Dialogue (SpeD), Bucharest, Romania, 2021, pp. 167-171

work page 2021

[7] [7]

Infant crying detection in real-world environments,

X. Yao, M. Micheletti, M. Johnson, E. Thomaz, and K. de Barbaro, "Infant crying detection in real-world environments," in Proc. of IEEE International Conference on Acoustics, Speech and Signal Processing, Singapore, 2022, pp. 131-135

work page 2022

[8] [8]

Is your baby fine at home? Baby cry sound detection in domestic environments,

T. Khandelwal, R. K. Das, and E. S. Chng, "Is your baby fine at home? Baby cry sound detection in domestic environments," in Proc. of Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Chiang Mai, Thailand, 2022, pp. 275-280

work page 2022