Pith. sign in

REVIEW 3 major objections 4 minor 59 references

RBA-FE: A Robust Brain-Inspired Audio Feature Extractor for Depression Diagnosis

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing LSTM gates with adaptive-rate spiking neurons yields a noise-tolerant audio classifier that reports 89.74% accuracy on MODMA and the best F1 among compared audio-only DAIC-WOZ baselines.

desk verdict Interesting spiking-neuron twist, but the headline accuracy is likely inflated by speaker leakage in the splits. read the letter →

arxiv 2506.07118 v1 pith:Q7WXHTNR submitted 2025-06-08 cs.SD cs.AIeess.AS

Yu-Xuan Wu , Ziyan Huang , Bin Hu , Zhi-Hong Guan This is my paper
classification cs.SDcs.AIeess.AS
keywords DepressiondiagnosisAudiofeatureextractionSpikingneuralnetworksLeakyintegrate-and-fireAdaptivethresholdNoiserobustnessSpeechanalysisAttentionmechanism
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 depression can be screened from short voice recordings in noisy, home-style conditions by a hierarchical audio feature extractor whose neuron model is noise-adaptive. The proposed RBA-FE pipeline takes six acoustic features, processes them through a temporal CNN, multi-head self-attention, and a bidirectional LSTM, and replaces the LSTM's sigmoid gates with a new spiking neuron, ARSLIF, whose firing threshold adapts to input statistics. The authors report state-of-the-art results on the MODMA dataset (89.74% accuracy, 0.8750 F1), lower regression error on AVEC2014 than several published baselines, and the best F1 among compared audio-only methods on DAIC-WOZ (0.6808), with the ARSLIF module outperforming sigmoid, soft-threshold, and standard LIF gates under injected pink, blue, and purple noise. If the claims hold, a fast, objective, equipment-light depression screener becomes much more plausible, and an adaptive spiking mechanism is added to the toolbox for noise-tolerant speech analysis.

What carries the argument

The load-bearing object is the ARSLIF neuron model, an adaptive-rate smooth leaky integrate-and-fire unit. Like a standard LIF neuron it accumulates membrane potential and fires on threshold crossing, but instead of a fixed threshold it updates $V_{\mathrm{th}}$ toward a target firing rate: the threshold rises when the firing-rate error $r_e = \alpha(r(t)-f) + \beta(r(t)-f_{\mathrm{adapt}}(t))$ is positive, with $f_{\mathrm{adapt}}$ switching between a high target rate and a low one depending on whether the input is above threshold. At steady state the firing rate becomes a weighted average of $f$ and $f_{\mathrm{adapt}}$, so $\alpha/(\alpha+\beta)$ sets the balance between stability and adaptability. The ARSLIF unit replaces sigmoid gates in the Bi-LSTM, and the paper's argument is that its threshold adaptation suppresses noise-induced spikes while preserving sparse, salient pulses; a smoothing operation is also added to stabilize training, and the heatmap comparison against sigmoid is offered as evidence.

What would settle it

A decisive test: re-run the MODMA classification under strict speaker-disjoint cross-validation, keeping every segment of each participant in one split; if the accuracy falls far below the reported 89.74%, the headline result was carried by speaker identity rather than depression.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a leaky integrate-and-fire neuron can be made noise-robust by making its activation threshold adaptive rather than fixed: ARSLIF raises its threshold when high firing would otherwise be driven by noise, and lowers it to preserve salient weak features, mimicking the brain's retuning of cellular signal selectivity under attention. Embedded as the gate activation of a Bi-LSTM inside a T-CNN–attention hierarchy, it yields a depression-audio classifier that the authors report as reaching 0.8750 precision, 0.8974 accuracy, 0.8750 recall, and 0.8750 F1 on MODMA, an F1 of 0.6808 on DAIC-WOZ, RMSE and MAE of 8.8310 on their merged AVEC2014 setup, and better error rates than sigmoid, soft thresholding, and standard LIF gates under -20 to +20 dB pink, blue, and purple noise. The paper also derives that the adaptive threshold reduces the power of noise-induced spiking, giving $\mathrm{SNR}_{\mathrm{ARSLIF}} > \mathrm{SNR}_{\mathrm{LIF}}$, and reports that the firing pattern on depressive audio differs from healthy audio, which it offers as a brain-inspired interpretability result.

Load-bearing premise

The reported accuracy numbers assume the same speaker's audio never appears in both the training and test splits, so the model cannot learn to identify people instead of identifying depression.

Editorial extensions

If this is right

  • On the MODMA dataset, RBA-FE reports 0.8974 accuracy and 0.8750 F1, higher than the DenseNet, EfficientNet, and SENet baselines in the paper's comparison.
  • On AVEC2014's merged regression task, RBA-FE reports RMSE 8.8310, lower than the compared baselines, with the authors noting that high-risk (high BDI) cases are emphasized.
  • On DAIC-WOZ, the audio-only RBA-FE reports F1 0.6808, exceeding DEPA, SIDD, DepAudioNet, and FRAUG in the paper's comparison.
  • Ablation results on AVEC2014 show removing multi-head attention raises MAE to 9.1419 and removing Bi-LSTM raises it to 11.2401, indicating that the hierarchical ordering is load-bearing.
  • The adaptive-threshold mechanism is tunable: the ratio $\alpha/(\alpha+\beta)$ trades stable firing for fast adaptation, and the paper selects $\alpha=0.6$, $\beta=0.4$ as the balance point.

Reading between the lines

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

  • Beyond the paper: if ARSLIF's threshold adaptation is the active noise filter, the same gate replacement should transfer to other speech-based health screenings or any noisy temporal classification, and that transfer is a direct test of the mechanism rather than of the depression pipeline.
  • Beyond the paper: the SNR argument assumes Gaussian white noise, but real home environments contain babble, reverberation, and colored transients; testing with actual room recordings or non-stationary noise bursts would show whether the threshold adaptation holds outside the synthetic pink/blue/purple-noise setup.
  • Beyond the paper: the reported abnormal firing pattern on depressive audio hints at using ARSLIF spike statistics as a biomarker; comparing firing-rate trajectories against clinician-rated symptom dimensions would test whether the pattern is depression-specific or merely arousal-related.
  • Beyond the paper: because the MODMA split is described as stratified on audio IDs rather than on confirmed participant independence, a speaker-disjoint replication is the decisive check that the 89.74% figure reflects depression detection rather than voice identity.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes RBA-FE, a hierarchical audio feature extractor for depression diagnosis that combines T-CNN, multi-head attention, Bi-LSTM, and a novel adaptive-rate smooth leaky integrate-and-fire (ARSLIF) neuron model. The ARSLIF model is claimed to emulate brain-like 'retuning of cellular signal selectivity' via an adaptive firing threshold, yielding improved noise robustness. The authors report state-of-the-art classification accuracy on the MODMA dataset (0.8974), lower RMSE on AVEC2014 (8.8310), and competitive F1 on DAIC-WOZ (0.6808), alongside noise-robustness and ablation experiments. The central claims are that RBA-FE outperforms prior audio-based depression diagnosis models and that ARSLIF provides a theoretically grounded SNR improvement over standard LIF.

Significance. If the empirical results were reproduced under properly controlled, subject-disjoint experimental protocols and the SNR advantage were rigorously derived, the work would offer a practically useful noise-tolerant speech-based depression screener and a biologically motivated spiking mechanism of modest interest. The paper includes several strengths: it evaluates on three public datasets, compares against multiple baselines, runs ablations, tests robustness under three types of synthetic noise, and benchmarks several adaptive LIF variants. However, the load-bearing evidence is currently undermined by unresolved data-split issues, an undemonstrated theoretical proof, and a lack of statistical reliability in the reported metrics.

major comments (3)
  1. [IV.G.1 and IV.A] The MODMA experiment splits '80% of all audio IDs' into training and 20% into testing without establishing that the same participant's segments never appear in both sets. Because the depression label is constant per participant and each participant contributes many audio segments, a random split on audio IDs almost certainly places the same speaker in both training and test, allowing the model to exploit speaker identity rather than depression-specific cues. The AVEC2014 preparation in IV.A is similarly problematic: the paper merges the original training, development, and test sets and averages BDI scores per participant 'within both the training and test sets,' without reporting whether any participant appears in both sets. Since the central empirical claim of state-of-the-art performance rests on these comparisons, the authors must provide subject-disjoint splits (or a detailed overlap analysis) and report results under such splits for Tables II, V, and VI.
  2. [III.C.2, Eqs. (12)-(13)] The claimed proof of improved SNR is not a derivation. Equation (12) defines the output noise power as the integral of η(t) minus a threshold-correction term, but ΔV_th(t) is never expressed in terms of the spiking dynamics or the noise process; the subsequent assertion that ΔV_th 'increases with noise' and hence reduces P_noise is the conclusion restated as an assumption. Moreover, the steady-state analysis leading to Eq. (11) incorrectly states that convergence of V_th implies the firing rate 'reaches zero,' whereas Eq. (10) implies only that the rate error vanishes and gives the non-zero equilibrium in Eq. (11). The theoretical advantage of ARSLIF over LIF therefore remains unsupported and needs a rigorous derivation or an explicitly stated empirical basis.
  3. [IV (Tables II, V, VI; Fig. 7)] All experimental claims are reported without error bars, confidence intervals, or repeated runs. Given the small number of subjects (52 in MODMA) and the large number of tunable hyperparameters in Table I, single-run metrics are insufficient to support the stated state-of-the-art performance. The comparisons to prior work also do not control for feature sets, input representations, or training budgets, so the differences in Tables II and VI may not be attributable to the proposed architecture. The paper should report mean and variance over multiple random seeds and, where possible, statistical significance tests.
minor comments (4)
  1. [IV.D, Table II] In the paragraph following Table II, the model is referred to as 'RBE-FA' instead of 'RBA-FE'; this typo appears multiple times and should be corrected.
  2. [I and IV.B] The introduction states that the model uses 'six acoustic features' and names 'MFCC, Pitch, Jitter, CQL Cepstrum,' which is a list of only four features. Later, in IV.B, the six features are described as MFCCs, first-order and second-order MFCC deltas, CQT, pitches, and jitter. The manuscript should reconcile the feature nomenclature (CQL vs. CQT) and consistently list all six features.
  3. [IV.E, Table III] Table III reports F1 scores for different adaptive LIF models but does not state which dataset is used; the same F1 value (68.08%) also appears for RBA-FE on DAIC-WOZ in Table VI, which may be coincidental or an unnoticed duplication. The authors should clarify the experimental setup for Table III.
  4. [IV.G.1, Fig. 8] The ROC curve caption says the blue line represents a random classifier and the orange curve represents the model, but the figure itself is not referenced in the text with these colors; adding an explicit legend or matching description would improve clarity.

Circularity Check

2 steps flagged · score 6.0 of 10

The claimed SNR advantage of ARSLIF is asserted by the definition of P_ARSLIF_noise (Eq. 12) rather than derived, and the AVEC2014 comparison uses hyperparameters tuned on the same data; MODMA/DAIC-WOZ results are independent, so circularity is partial.

  1. self definitional [Section III.C.2, Eqs. (12)-(13)]
    "For the ARSLIF model, the adaptive threshold V_th(t) increases with a raising noise by (9), with the goal of filtering noise-induced spikes. So the power of noise-induced spiking in ARSLIF is reduced because the threshold increases with noise: P_ARSLIF_noise = ∫_0^T (η(t) − (1/τ_adapt) ΔV_th(t))^2 dt (12). Since ΔV_th(t) increases with noise, this term reduces the overall power of noise-induced spiking, yielding an improved SNR. More specifically, one has SNR_ARSLIF = P_signal / P_ARSLIF_noise > P_signal / P_noise = SNR_LIF (13)."

    Equation (12) is not derived from the ARSLIF dynamics (3)-(9); it simply defines the 'noise-induced spiking power' as the residual after subtracting a threshold-adaptation term. The inequality (13) is then read off from that definition. To prove the inequality one would need a bound such as ∫ η ΔV_th > (1/(2 τ_adapt)) ∫ (ΔV_th)^2, but no such estimate is given. The conclusion that the adaptive threshold 'reduces the overall power' is therefore an assumption built into the definition of P_ARSLIF_noise, not a consequence of the spiking equations. The claimed theoretical advantage of ARSLIF over LIF is effectively the statement of its own construction.

  2. fitted input called prediction [Section IV.C, 'Choosing Adaptive Threshold in Training' (and Table II)]
    "Here, we chose a 60% target rate error ratio: α= 0.6 and β= 0.4, which experimentally proves to effectively balance volatility and stability for A VEC2014."

    α and β determine the adaptive threshold update (Eq. 9) and hence the model's firing behaviour. They are chosen by inspecting threshold curves and performance on AVEC2014, and the same dataset is then used in Table II to report RBA-FE's RMSE/MAE as state-of-the-art. The AVEC2014 comparison is therefore partly an in-sample report of the tuning criterion rather than an independent prediction. This is standard hyperparameter selection, but it becomes circular when the tuned configuration is presented as evidence that ARSLIF outperforms alternatives on the same data. The MODMA and DAIC-WOZ experiments are not affected by this particular fit, which keeps the circularity partial.

full rationale

The paper's central derivation chain was examined. The main circular step is the theoretical noise-robustness analysis in Sec. III.C.2: Eq. (12) defines the ARSLIF noise power as the residual after subtracting the threshold-adaptation term, and Eq. (13) asserts the SNR improvement without deriving the required correlation inequality; the advantage is built into the definition. A second, milder circularity is the selection of α=0.6, β=0.4 on AVEC2014 data in Sec. IV.C followed by the AVEC2014 comparison in Table II; the fitted hyperparameters are presented as evidence for the same dataset. However, the MODMA and DAIC-WOZ experiments (Secs. IV.G.1-IV.G.2) use fixed hyperparameters and provide independent empirical support, as do the noise-injection comparisons in Fig. 7 (though those curves also share the AVEC2014 tuning). No load-bearing self-citations were found: refs [16], [17], [22] are standard neuroscience sources, and the authors' own ref [19] is only peripheral support for hybrid networks. No uniqueness theorem is imported, and no ansatz is smuggled in via citation. The acknowledged limitation about global max pooling losing intra-frame information (Sec. IV.D) is an honest statement, not a circular step. The speaker-split concern for MODMA and the merged AVEC2014 splits is a serious validity risk but belongs to experimental correctness rather than circularity under the rubric. Overall, the theoretical claim reduces by construction while the empirical claims retain independent content, giving a partial circularity score of 6.

Assumptions & free parameters 9 free parameters · 7 assumptions · 1 invented entities

The central empirical results rest on a pipeline of manually tuned parameters, borrowed audio features, and a new neuron model whose theoretical advantage is asserted. The robustness argument in Sec III.C.2 is not a derivation from first principles; it constructs the noise-reduced output by definition. The model also relies on standard LIF, attention, LSTM, and feature-engineering machinery as unproved background.

free parameters (9)
  • Alpha (target rate error ratio) = 0.6 (AVEC2014); 0.6 (MODMA/DAIC-WOZ)
    Chosen by experiments on AVEC2014 (Sec IV.C, Fig 6) to balance stability and adaptability; carried to other datasets.
  • Beta (adaptive rate error ratio) = 0.4
    Complements alpha=0.6 so alpha+beta=1; selected via the same AVEC2014 tuning.
  • Target firing rate f = 0.6
    Manually set for all datasets (Table I, Sections IV.B, IV.G.1, IV.G.2).
  • Initial activation threshold V_th = 0.8 (AVEC2014), 0.6 (MODMA and DAIC-WOZ)
    Manually tuned per dataset and documented in Table I and Section IV.G.
  • Active target rate f_active = 0.99
    Set in Table I as the target firing rate when input exceeds threshold.
  • Rest target rate f_rest = 0.01
    Set in Table I as the target firing rate when input is below threshold.
  • Threshold adaptation time constant tau_adapt = 1000
    Set in Table I; controls the speed of threshold adjustment.
  • Learning rate = 0.0005 (AVEC2014), 0.001 (MODMA/DAIC-WOZ)
    Adam optimizer setting, manually chosen and stated per experiment.
  • Batch size = 96
    Set for efficiency in all experiments.
assumptions (7)
  • domain assumption Standard LIF membrane potential and spike generation equations (Eq 1-2) describe neural spiking sufficiently for noise filtering.
    Adopted from [16],[17]; not proved in this paper but standard in the SNN literature.
  • domain assumption Input speech contains additive Gaussian white noise with zero mean and power sigma^2 * T.
    Standard noise model used in Sec III.C.2; not validated for real environmental noise.
  • ad hoc to paper The adaptive threshold update (Eq 6) using rate error (Eq 5) makes V_th increase with noise, which then cancels noise in spike output.
    This is the paper's own mechanism; Eq (12) assumes the cancellation without derivation from the actual spiking dynamics.
  • ad hoc to paper The average firing rate r(t) of all neurons in the ARSLIF layer is the appropriate control variable for adjusting each neuron's threshold.
    Eq (4) defines r(t) as a global average; the paper does not justify its use per neuron.
  • ad hoc to paper Replacing the sigmoid gates of Bi-LSTM with ARSLIF preserves or improves learning (Eq 14).
    This substitution is proposed by the authors; no formal convergence or stability analysis is given.
  • domain assumption The six handcrafted acoustic features (MFCC, deltas, CQT, pitch, jitter) carry sufficient information for depression diagnosis.
    Based on prior studies [24]-[26],[41]; not newly validated here.
  • ad hoc to paper The brain mechanism of 'retuning of cellular signal selectivity' maps to an adaptive firing threshold that rises for non-preferred stimuli.
    The paper draws this analogy from [22] but the mapping to ARSLIF's threshold rule is the authors' own.
invented entities (1)
  • ARSLIF neuron model (adaptive rate smooth leaky integrate-and-fire)
    purpose: Replaces sigmoid gates in Bi-LSTM to filter noise and adapt thresholds to input statistics.
    It is a new computational construct; the paper provides no independent biological or empirical evidence beyond its own experiments that the 'cell selectivity' analogy is valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RBA-FE: A Robust Brain-Inspired Audio Feature Extractor for Depression Diagnosis." pith.science (2026). https://pith.science/paper/Q7WXHTNR

@misc{pith2026250607118,
  author       = {Pith},
  title        = {Pith review of: RBA-FE: A Robust Brain-Inspired Audio Feature Extractor for Depression Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q7WXHTNR}},
  note         = {Machine review of arXiv:2506.07118}
}
read the original abstract

This article proposes a robust brain-inspired audio feature extractor (RBA-FE) model for depression diagnosis, using an improved hierarchical network architecture. Most deep learning models achieve state-of-the-art performance for image-based diagnostic tasks, ignoring the counterpart audio features. In order to tailor the noise challenge, RBA-FE leverages six acoustic features extracted from the raw audio, capturing both spatial characteristics and temporal dependencies. This hybrid attribute helps alleviate the precision limitation in audio feature extraction within other learning models like deep residual shrinkage networks. To deal with the noise issues, our model incorporates an improved spiking neuron model, called adaptive rate smooth leaky integrate-and-fire (ARSLIF). The ARSLIF model emulates the mechanism of ``retuning of cellular signal selectivity" in the brain attention systems, which enhances the model robustness against environmental noises in audio data. Experimental results demonstrate that RBA-FE achieves state-of-the-art accuracy on the MODMA dataset, respectively with 0.8750, 0.8974, 0.8750 and 0.8750 in precision, accuracy, recall and F1 score. Extensive experiments on the AVEC2014 and DAIC-WOZ datasets both show enhancements in noise robustness. It is further indicated by comparison that the ARSLIF neuron model suggest the abnormal firing pattern within the feature extraction on depressive audio data, offering brain-inspired interpretability.

Figures

Figures reproduced from arXiv: 2506.07118 by the authors.

Figure 1
Figure 1. Diagram of the RBA-FE model. Initial inputs are the preprocessed six key features obtained from raw audio signals. These features go through [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The hierarchical network architecture for feature extraction in the RBA-FE model, with inspiration from the information processing in auditory pathway [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of data flow and tensor sizes in RBA-FE. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Illustration of the calculation of ARSLIF based on equations (3)-(8), [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparative Analysis of ARSLIF and Sigmoid Responses to Clean and Noisy Signals. (a) is the unperturbed input signal, (b) represents ARSLIF [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Changes of threshold under different target frequency error ratios [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: A comprehensive perspective for comparing the performance of four different activation functions against ARSLIF [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: ROC curves of comparative models on the testing MODMA dataset [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 57 canonical work pages

  1. [1]

    Depression,

    World Health Organization, “Depression,” n.d. [Online]. Available: https: //www.who.int/zh/health-topics/depression

    work page

  2. [2]

    Depression,

    National Institutes of Health, “Depression,” n.d. [Online]. Available: https: //www.nimh.nih.gov/sites/default/files/documents/health/publications

    work page

  3. [3]

    The SCL-90-R, Brief Symptom Inventory, and Matching Clinical Rating Scales,

    L. R. Derogatis and K. L. Savitz, “The SCL-90-R, Brief Symptom Inventory, and Matching Clinical Rating Scales,” in M. E. Maruish (Ed.), The use of psychological testing for treatment planning and outcomes assessment, Lawrence Erlbaum Associates Publishers, 1999, pp. 679-724

    work page 1999

  4. [4]

    Self-rating depression scale in an outpatient clinic: further validation of the SDS,

    W. W. Zung, C. B. Richards, and M. J. Short, “Self-rating depression scale in an outpatient clinic: further validation of the SDS,” Archives of General Psychiatry, vol. 13, no. 6, pp. 508-515, 1965

    work page 1965

  5. [5]

    Interpretation of Depression Detection Models via Feature Selection Methods,

    S. Alghowinem, T. Gedeon, R. Goecke, J. F. Cohn, and G. Parker, “Interpretation of Depression Detection Models via Feature Selection Methods,” in IEEE Trans. Affect. Comput., vol. 14, no. 1, pp. 133-152, 1 2023

    work page 2023

  6. [6]

    WavDepressionNet: Automatic Depression Level Prediction Via Raw Speech Signals,

    M. Niu, J. Tao, Y . Li, Y . Qin, and Y . Li, “WavDepressionNet: Automatic Depression Level Prediction Via Raw Speech Signals,” in IEEE Trans. Affect. Comput., Early Access

    work page

  7. [7]

    SpeechFormer++: A Hierarchical Efficient Framework for Paralinguistic Speech Processing,

    W. Chen, X. Xing, X. Xu, J. Pang, and L. Du, “SpeechFormer++: A Hierarchical Efficient Framework for Paralinguistic Speech Processing,” in IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 31, pp. 775-788, 2023

    work page 2023

  8. [8]

    Generalized Two-Stage Rank Regression Framework for Depression Score Prediction from Speech,

    N. Cummins, V . Sethu, J. Epps, J. R. Williamson, T. F. Quatieri, and J. Krajewski, “Generalized Two-Stage Rank Regression Framework for Depression Score Prediction from Speech,” in IEEE Trans. Affect. Comput., vol. 11, no. 2, pp. 272-283, 2020

    work page 2020

Show all 59 references
  1. [9]

    High-Density Electroencephalography and Speech Signal Based Deep Framework for Clinical Depression Diagnosis,

    A. Qayyum, I. Razzak, M. Tanveer, M. Mazher, and B. Alhaqbani, “High-Density Electroencephalography and Speech Signal Based Deep Framework for Clinical Depression Diagnosis,” in IEEE/ACM Trans. Comput. Biol. Bioinf., vol. 20, no. 4, pp. 2587-2597, 2023

    2023

  2. [10]

    Multimodal Spatiotem- poral Representation for Automatic Depression Level Detection,

    M. Niu, J. Tao, B. Liu, J. Huang, and Z. Lian, “Multimodal Spatiotem- poral Representation for Automatic Depression Level Detection,” in IEEE Trans. Affect. Comput., vol. 14, no. 1, pp. 294-307, 2023

    2023

  3. [11]

    Prediction of Depression Severity Based on the Prosodic and Semantic Features with Bidirectional LSTM and Time Distributed CNN,

    K. Mao, W. Zhang, D. B. Wang, et al., “Prediction of Depression Severity Based on the Prosodic and Semantic Features with Bidirectional LSTM and Time Distributed CNN,” in IEEE Trans. Affect. Comput., Early Access

  4. [12]

    Frontier report,

    UNEP, “Frontier report,” 2022. [Online]. Available: https://www.unep. org/zh-hans/events/publication%20launch/2022qianyanbaogaofabu

    2022

  5. [14]

    Deep residual shrinkage networks for fault diagnosis,

    M. Zhao, S. Zhong, and X. Fu, “Deep residual shrinkage networks for fault diagnosis,” in IEEE Trans. Ind. Informatics, vol. 16, no. 7, pp. 4681- 4690, 2020

    2020

  6. [15]

    Bioacoustic features of depression: A review,

    S. A. Almaghrabi, S. R. Clark, and M. Baumert, “Bioacoustic features of depression: A review,” in Biomedical Signal Processing and Control, vol. 85, 2023, p. 105020

    2023

  7. [16]

    Principles of Neural Science,

    E. R. Kandel, J. H. Schwartz, T. M. Jessell, S. A. Siegelbaum, A. J. Hudspeth, and S. Mack, “Principles of Neural Science,” Fifth Edition, McGraw Hill, 2014

    2014

  8. [17]

    Action Potential Parame- ters and Spiking Behavior of Cortical Neurons: A Statistical Analysis for Designing Spiking Neural Networks,

    A. Chakraborty, S. Panda, and S. Chakrabarti, “Action Potential Parame- ters and Spiking Behavior of Cortical Neurons: A Statistical Analysis for Designing Spiking Neural Networks,” in IEEE Trans. Cogn. Dev. Syst., vol. 15, no. 2, pp. 808-818, 2023

    2023

  9. [18]

    Towards artificial general intelligence with hybrid Tianjic chip architecture,

    J. Pei, L. Deng, S. Song, et al. “Towards artificial general intelligence with hybrid Tianjic chip architecture,” in Nature, vol. 572, pp. 106-111, 2019

    2019

  10. [19]

    Multistability of delayed hybrid impulsive neural networks with application to associative memories,

    B. Hu, Z.-H. Guan, G. Chen, and F. L. Lewis. “Multistability of delayed hybrid impulsive neural networks with application to associative memories,” in IEEE Trans. Neura. Net. Learn. Syst., vol. 30, no. 5, pp. 1537-1555, 2019

    2019

  11. [20]

    A comprehensive analysis on adversarial robustness of spiking neural networks,

    Sharmin, S., Panda, P., Sarwar, S. S., Lee, C., Ponghiran, W., and Roy, K., “A comprehensive analysis on adversarial robustness of spiking neural networks,” in 2019 IEEE International Joint Conference on Neural Networks (IJCNN), pp. 1-8

    2019

  12. [21]

    Long short-term memory and learning-to-learn in networks of spiking neurons,

    G. Bellec, D. Salaj, A. Subramoney, R. Legenstein, and W. Maass, “Long short-term memory and learning-to-learn in networks of spiking neurons,” in NeurIPS 2018

    2018

  13. [22]

    Attention increases sensitivity of V4 neurons,

    J. H. Reynolds, T. Pasternak, and R. Desimone, “Attention increases sensitivity of V4 neurons,” Neuron, vol. 26, no. 3, pp. 703-714, 2000

    2000

  14. [23]

    A review of depression and suicide risk assessment using speech analysis,

    N. Cummins, S. Scherer, J. Krajewski, S. Schnieder, J. Epps, and T. F. Quatieri, “A review of depression and suicide risk assessment using speech analysis,” Speech Communication, vol. 71, pp. 10-49, 2015

    2015

  15. [24]

    V ocal indicators of mood change in depression,

    H. Ellgring and K. R. Scherer, “V ocal indicators of mood change in depression,” Journal of Nonverbal Behavior, vol. 20, pp. 83-110, 1996

    1996

  16. [25]

    Major depressive disorder discrimination using vocal acoustic features,

    T. Taguchi et al., “Major depressive disorder discrimination using vocal acoustic features,” Journal of Affective Disorders, vol. 225, pp. 214-220, 2018. 14

    2018

  17. [26]

    Investigation of vocal jitter and glottal flow spectrum as possible cues for depression and near-term suicidal risk,

    A. Ozdas, R. G. Shiavi, S. E. Silverman, M. K. Silverman, and D. M. Wilkes, “Investigation of vocal jitter and glottal flow spectrum as possible cues for depression and near-term suicidal risk,” in IEEE Trans. Biomed. Eng., vol. 51, no. 9, pp. 1530-1540, 2004

    2004

  18. [27]

    Squeeze-and-excitation networks

    J. Hu, L. Shen, and G. Sun. “Squeeze-and-excitation networks”. in: Proceedings of the IEEE conference on computer vision and pattern recognition. 2018, pp. 7132-7141

    2018

  19. [28]

    Automated depression analysis using convolutional neural networks from speech,

    L. He and C. Cao, “Automated depression analysis using convolutional neural networks from speech,” In: J. Biomed. Inform., vol. 83, pp. 103- 111, 2018

    2018

  20. [29]

    Speech emotion recognition using deep 1D & 2D CNN LSTM networks,

    J. Zhao, X. Mao, and L. Chen, “Speech emotion recognition using deep 1D & 2D CNN LSTM networks,” Biomed. Signal Process. Control, vol. 47, pp. 312-323, 2019

    2019

  21. [30]

    Towards robust deep neural networks for affect and depression recognition from speech,

    A. Othmani et al., “Towards robust deep neural networks for affect and depression recognition from speech,” in Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10-15, 2021

    2021

  22. [31]

    Noise and reverberation effects on depression detection from speech,

    V . Mitra, A. Tsiartas, and E. Shriberg, “Noise and reverberation effects on depression detection from speech,” in 2016 IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), 2016, pp. 5795-5799

    2016

  23. [32]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems. Curran Associates Inc., 2017, pp. 6000-6010

    2017

  24. [33]

    Bidirectional LSTM-CRF models for sequence tagging,

    Z. Huang, W. Xu, and K. Yu, “Bidirectional LSTM-CRF models for sequence tagging,” in arXiv preprint arXiv:1508.01991, 2015 [cs.CL], [Online]. Available: https://doi.org/10.48550/arXiv.1508.01991

    2015 doi arXiv

  25. [34]

    Recurrent neural network regularization,

    W. Zaremba, I. Sutskever, and O. Vinyals, “Recurrent neural network regularization,” in arXiv preprint arXiv:1409.2329, 2014 [cs.NE], [On- line]. Available: https://doi.org/10.48550/arXiv.1409.2329

    2014 doi arXiv

  26. [35]

    A theoretically grounded application of dropout in recurrent neural networks,

    Y . Gal and Z. Ghahramani, “A theoretically grounded application of dropout in recurrent neural networks,” in Advances in Neural Information Processing Systems (NIPS).vol. 29, pp. 30-38, 2016,

    2016

  27. [36]

    A VEC 2014: 3D Dimensional Affect and Depression Recognition Challenge,

    M. Valstar et al., “A VEC 2014: 3D Dimensional Affect and Depression Recognition Challenge,” in Proceedings of the 4th International Workshop on Audio/Visual Emotion Challenge (A VEC ’14), 2014, pp. 3-10

    2014

  28. [37]

    A multi-modal open dataset for mental- disorder analysis,

    H. Cai, Z. Yuan, Y . Gao, et al., “A multi-modal open dataset for mental- disorder analysis,” in Scientific Data, vol. 9, no. 1, pp. 178, 2022

    2022

  29. [38]

    V ocal acoustic biomarkers of depression severity and treatment response,

    J. C. Mundt, A. P. V ogel, D. E. Feltner, and W. R. Lenderking, “V ocal acoustic biomarkers of depression severity and treatment response,” Biological Psychiatry, vol. 72, no. 7, pp. 580-587, 2012

    2012

  30. [39]

    Depression Estimation Using Audiovisual Features and Fisher Vector Encoding,

    V . Jain, J. L. Crowley, A. K. Dey, and A. Lux, “Depression Estimation Using Audiovisual Features and Fisher Vector Encoding,” in Proceedings of the 4th International Workshop on Audio/Visual Emotion Challenge (A VEC ’14), Association for Computing Machinery, New York, NY , US...

    2014

  31. [40]

    Artificial Intelligent System for Automatic Depression Level Analysis Through Visual and V ocal Expressions,

    A. Jan, H. Meng, Y . F. B. A. Gaus, and F. Zhang, “Artificial Intelligent System for Automatic Depression Level Analysis Through Visual and V ocal Expressions,” in IEEE Trans. Cogn. Dev. Syst., vol. 10, no. 3, pp. 668-680, 2018

    2018

  32. [41]

    V ocal-source biomarkers for depression: A link to psychomotor activity,

    T. F. Quatieri and N. Malyska, “V ocal-source biomarkers for depression: A link to psychomotor activity,” in Interspeech, vol. 2012, no. 2, 2012, pp. 1059-1062

    2012

  33. [42]

    Automatic Depression Level Detection viaℓ p-Norm Pooling,

    M. Niu, J. Tao, B. Liu, and C. Fan, “Automatic Depression Level Detection viaℓ p-Norm Pooling,” in Interspeech, 2019

    2019

  34. [43]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition, 2017, pp. 4700-4708

    2017

  35. [44]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” in International Conference on Machine Learning, PMLR, 2019, pp. 6105-6114

    2019

  36. [45]

    EfficientNetV2: Smaller models and faster training,

    M. Tan and Q. Le, “EfficientNetV2: Smaller models and faster training,” in International Conference on Machine Learning, PMLR, 2021, pp. 10096-10106

    2021

  37. [46]

    Uncertainty Estimation for Heatmap- Based Landmark Localization,

    L. A. Schobs, A. J. Swift, H. Lu, “Uncertainty Estimation for Heatmap- Based Landmark Localization,” IEEE Trans. Med. Imag. vol. 42, no. 4, pp. 1021-1034, 2023

    2023

  38. [47]

    Detection of clinical depression in adoles- cents’ speech during family interactions

    Low, Lu-Shih Alex, et al. “Detection of clinical depression in adoles- cents’ speech during family interactions.” IEEE Transactions on Biomed- ical Engineering, vol. 58, no. 3, pp. 574-586, 2010

    2010

  39. [48]

    A review of depression and suicide risk assessment using speech analysis

    Cummins, Nicholas, et al. “A review of depression and suicide risk assessment using speech analysis.” Speech Communication, vol. 71, pp. 10-49, 2015

    2015

  40. [49]

    V ocal-source biomarkers for depression: A link to psychomotor activity

    Quatieri, Thomas F., and Nicole Malyska. “V ocal-source biomarkers for depression: A link to psychomotor activity.” Interspeech 2012, Portland, Oregon, 2012, pp. 1059-1062

    2012

  41. [50]

    Investigation of vocal jitter and glottal flow spectrum as possible cues for depression and near-term suicidal risk

    Ozdas, A., et al. “Investigation of vocal jitter and glottal flow spectrum as possible cues for depression and near-term suicidal risk.” IEEE Transactions on Biomedical Engineering, vol. 51, no. 9, pp. 1530-1540, 2004

    2004

  42. [51]

    Exploring harmonic structure in the analysis of emotional speech

    Almaghrabi, Rawan, et al. “Exploring harmonic structure in the analysis of emotional speech.” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2016, pp. 5195-5199

    2016

  43. [52]

    3-10, 2016

    Michel Valstar, Jonathan Gratch et al., ”Avec 2016: Depression mood and emotion recognition workshop and challenge”, Proceedings of the 6th international workshop on audio/visual emotion challenge, pp. 3-10, 2016

    2016

  44. [53]

    Depa: Self-supervised audio embedding for depression detection

    Zhang, Pingyue, et al. “Depa: Self-supervised audio embedding for depression detection.” Proceedings of the 29th ACM international con- ference on multimedia. 2021

    2021

  45. [54]

    Lishi Zuo, Man-Wai Mak, Avoiding dominance of speaker features in speech-based depression detection, Pattern Recognition Letters, V olume 173, 2023, pp. 50-56

    2023

  46. [55]

    the 6th International Workshop on Audio/Visual Emotion Challenge, 2016, pp

    Xingchen Ma, Hongyu Yang, Qiang Chen, Di Huang, Yunhong Wang, Depaudionet: An efficient deep model for audio based depression classifi- cation, in: Proc. the 6th International Workshop on Audio/Visual Emotion Challenge, 2016, pp. 35–42

    2016

  47. [56]

    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp

    Vijay Ravi, Jinhan Wang, Jonathan Flint, Abeer Alwan, Fraug: A Frame Rate Based Data Augmentation Method for Depression Detection from Speech Signals, in: Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6267–6271

    2022

  48. [57]

    Subdivisions of auditory cortex and processing streams in primates,

    J. H. Kaas and T. A. Hackett, “Subdivisions of auditory cortex and processing streams in primates,” Proc. Natl. Acad. Sci., vol. 97, no. 22, pp. 11793-11799, 2000

    2000

  49. [58]

    Dissecting neural computations in the human auditory pathway using deep neural networks for speech,

    Y . Li, G. K. Anumanchipalli, A. Mohamed, P. Chen, L. H. Carney, J. Lu, J. Wu, and E. F. Chang, “Dissecting neural computations in the human auditory pathway using deep neural networks for speech,” Nat. Neurosci., vol. 26, pp. 1484–1494, 2023

    2023

  50. [59]

    Deep neural network models of sensory systems: windows onto the role of task constraints,

    A. J. E. Kell and J. H. McDermott, “Deep neural network models of sensory systems: windows onto the role of task constraints,” Curr. Opin. Neurobiol., vol. 55, pp. 121-132, 2019

    2019

  51. [60]

    Processing Streams in Auditory Cortex,

    J. P. Rauschecker and B. Tian, “Processing Streams in Auditory Cortex,” Neural Correlates of Auditory Cognition, Springer Handbook of Auditory Research, vol. 45, pp. 7–43, 2012

    2012

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.