Pith. sign in

REVIEW 4 major objections 5 minor 39 references

On the reliability of feature attribution methods for speech classification

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Standard feature attribution methods are unreliable on speech models, except when perturbations are word-aligned for word-based tasks.

desk verdict Useful empirical study of attribution reliability in speech, with a load-bearing assumption about inter-seed feature similarity that should be validated or the claims softened. read the letter →

arxiv 2505.16406 v1 pith:FLCRZSLN submitted 2025-05-22 cs.CL eess.AS

Gaofei Shen , Hosein Mohebbi , Arianna Bisazza , Afra Alishahi , Grzegorz Chrupa{l}a This is my paper · ORCID
classification cs.CLeess.AS
keywords speechprocessingfeatureattributionreliabilityinter-seedagreementinterpretabilityperturbationmethodswordalignmentintentclassification
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

Feature attribution methods claim to show which parts of an input drive a model's decision, but explanations that change between equally good retrained copies of the same model cannot be trusted. This paper tests four standard attribution methods on speech classifiers and measures reliability as the agreement between attributions from nine fine-tuning seeds. It finds that most methods are unreliable whether the input is a waveform, spectrogram, or CNN embedding, and that post-hoc aggregation of scores does not fix this. The exception is directly perturbing word-aligned segments on word-based intent classification, where agreement becomes high. The upshot is that off-the-shelf attribution methods should not be presumed reliable for speech, and that reliable explanation requires aligning perturbation with meaningful lexical units for tasks that depend on them.

What carries the argument

The central instrument is the inter-seed agreement (ISA) metric, a dynamic top-p overlap score that measures how much the top 20% of attribution indices from two models fine-tuned with different random seeds intersect; it operationalizes reliability as reproducibility under retraining. The word-aligned perturbation setting, where the input is masked at forced-alignment word boundaries, is the treatment that rescues reliability for lexical tasks, and it works by removing the model's freedom to rely on different redundant cues across seeds.

What would settle it

Perturb the same intent-classification inputs using random segments of the same average length as words instead of forced-aligned word boundaries; if the inter-seed agreement stays high, the reliability effect is due to segment duration rather than lexical alignment, and the paper's exception would need reinterpretation.

Watch

Extended reading notes

Core claim

The paper claims that when a speech classifier is fine-tuned multiple times from the same pre-trained checkpoint with different random seeds, standard feature attribution methods—Saliency, Integrated Gradients, LIME, and Feature Ablation—produce explanations that agree poorly across runs, even though the classifiers themselves agree on nearly all test inputs. Using a dynamic top-20% inter-seed agreement (ISA) score, the authors find that for waveform, spectrogram, and CNN-embedding input types, and at no, frame, and word-level aggregation, most method-task combinations fall near or below a randomly shuffled baseline. The one reliable configuration is directly perturbing word-aligned input segments, and only for the word-based intent classification subtasks; gender and speaker identification never reach acceptable reliability. The conclusion is that naively applying standard attribution methods to speech generally yields unreliable explanations, and that reliable attribution requires aligning perturbations with meaningful lexical units for tasks that depend on them.

Load-bearing premise

The approach assumes that differently seeded fine-tuned models that agree on predictions also rely on the same internal features, so any disagreement in their attribution maps is due to the attribution method rather than to genuinely different model behavior.

Editorial extensions

If this is right

  • Practitioners should not treat gradient-based or high-resolution perturbation attributions for speech as stable explanations, since most conditions fall near the random baseline.
  • Aggregating high-resolution attribution scores to frame or word level does not systematically improve reliability for non-word-based tasks.
  • Reliable explanations in speech are currently restricted to tasks whose decisions depend on lexical content, using direct word-level perturbation.
  • Reliability must be established before validity: without seed-invariant attributions, faithfulness comparisons are themselves unstable.
  • Speech-specific attribution methods are needed for acoustic-dependent tasks like speaker and gender identification.

Reading between the lines

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

  • The ISA protocol could be transferred to other continuous-input domains, such as medical time series or audio event detection, where retraining stability is similarly under-evaluated.
  • Because the assumption that equally accurate seeds use comparable features is untested for acoustic tasks, the low ISA on gender and speaker ID may partly reflect genuinely different feature usage, not attribution failure; separating these two causes would strengthen the conclusion.
  • A natural extension would zoom into phoneme-level alignment, especially for languages where words are not the meaningful unit, to see whether the reliable regime generalizes beyond lexical segmentation.
  • The finding suggests a design rule that may transfer to other modalities: choose attribution granularity to match the model's decision units rather than the input resolution.
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

4 major / 5 minor

Summary. The paper presents an empirical study of the reliability of four feature attribution methods (Saliency, Integrated Gradients, LIME, and Feature Ablation) for speech classification. Using wav2vec2-base fine-tuned with nine random seeds on gender/speaker identification and intent classification (with three subtasks), the authors define the inter-seed agreement (ISA) metric as the fraction of shared top-20% attribution indices between pairs of runs. They report that ISA is generally low for waveform, spectrogram, and CNN-embedding attributions, even after frame- or word-level aggregation, but becomes high for word-aligned perturbation on intent classification subtasks. The paper concludes that standard attribution methods are generally unreliable in speech, except for word-aligned perturbation for word-based tasks.

Significance. If the finding is supported, it is a valuable and actionable result for the speech interpretability community: it would caution against off-the-shelf attribution methods on wav2vec2-style models and identify a concrete configuration (word-level perturbation for word-based tasks) that yields reproducible explanations. The study is well-grounded in standard datasets (Common Voice, Fluent Speech Commands) and the authors provide code, which aids reproducibility. The choice of inter-seed agreement as a reliability measure is reasonable in principle, but the paper's central negative claim hinges on an unvalidated assumption about inter-seed feature equivalence, and the lack of statistical tests and a single backbone limit the strength of the generalization.

major comments (4)
  1. [Section 3, Eq. (1)] The interpretation of ISA as a measure of attribution reliability rests entirely on the assumption, stated in Section 3, that models fine-tuned from the same checkpoint with different seeds 'identify and employ comparable relevant features for a given input' when they achieve high accuracy. The evidence offered in Table 1 (accuracy and Fleiss's kappa) is at the output level and does not establish internal feature equivalence; in a high-dimensional, redundant input space such as speech, different seeds could plausibly reach identical decisions via different features, in which case low ISA would reflect model variation rather than attribution unreliability. This is the load-bearing premise of the paper's negative conclusion, and it is not directly validated. Please add an explicit test: for example, compute a representation-similarity metric (e.g., CKA) between the fine-tuned models at the layers used for attribution, or use a synthetic task with known ground-truth features to show that the ISA gap between methods is driven by the attribution method rather than by inter-seed model differences.
  2. [Section 3.2 and Section 4] The paper fixes the top-p threshold at 20% and defines 'acceptable' reliability only by visual inspection of boxplots; no formal threshold, confidence interval, or significance test is reported. For instance, the statement in Section 5 that Integrated Gradients 'does not surpass 50% inter-seed agreement for most tasks' is a descriptive summary, and the claim that LIME is 'barely above the baseline' is not backed by a test of whether the distribution of ISA across the 36 seed pairs differs from the shuffled baseline. The central negative claim would be materially strengthened by reporting a permutation test against the random-shuffle baseline, confidence intervals on the medians, and a sensitivity analysis for the choice of p (e.g., 1%, 5%, 10%, 20%, 50%).
  3. [Section 3.1] All experiments use a single backbone, wav2vec2-base, and the abstract's conclusion that standard attribution methods are 'generally unreliable when applied to the speech domain' is broader than what the evidence supports. The paper should either temper the generalization to wav2vec2-base models (or to the tested conditions) or add at least one additional self-supervised backbone (e.g., HuBERT, WavLM) to test whether the reliability pattern transfers.
  4. [Section 4.3 and Section 5] The positive exception—word-aligned perturbation for intent classification—is the key result that prevents the paper from being entirely negative, but the mechanistic explanation given in Section 5 ('classification decisions for these tasks rely on specific words') is not directly tested. The lower ISA for the Location subtask is explained post hoc as due to 'redundant lexical cues,' but no supporting analysis is provided. I recommend adding a simple diagnostic: for the intent subtasks, compute whether the words identified by the word-aligned perturbation method correspond to task-specific lexical cues, or compare ISA for subsets of utterances with vs. without such cues. This would turn the explanation from a plausible conjecture into a tested claim.
minor comments (5)
  1. [Section 5] In the first paragraph, 'atrribution' should be corrected to 'attribution'.
  2. [Equation (1)] Equation (1) is not typeset cleanly in the manuscript; the product of N and the sum should be spaced clearly to avoid confusion.
  3. [Section 3.2] The value of p is said to be fixed at 20%, but the formula uses p as a variable; please clarify whether p is a proportion (0.2) or a percentage (20).
  4. [Figures 1-3] The red dotted lines are called 'randomly shuffled baseline' but the text does not explain how many shuffles were used or the variance of the baseline; please specify the procedure.
  5. [Section 3.1] The phrasing 'We can see that the classification accuracy is near perfect' is informal; also, the two sentences beginning 'If models' architecture...' are redundant with the presentation of Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical measurement study whose central claim summarizes an operationalized reliability metric applied under an explicit, non-circular assumption.

full rationale

The paper does not derive a fitted quantity and present it as a prediction. Its central claim is that standard feature attribution methods show poor inter-seed agreement (ISA) on speech models, except for word-aligned perturbation on word-based tasks. ISA is defined operationally in Eq. 1 as the overlap of top-20% attribution indices between models fine-tuned from the same checkpoint with different random seeds, and it is compared against a randomly shuffled baseline. No parameter is fitted to the data and then renamed as a finding, and no equation reduces to another by construction. The load-bearing assumption that different seeds 'identify and employ comparable relevant features' for high-accuracy models is explicit in Section 3 and is an empirical premise about model behavior, not a definitional identity. The authors attempt to support it with accuracy and Fleiss' kappa in Table 1; even if that support is debated, the conclusion would be invalidated by a false premise rather than be circular. Citations to prior work, including one co-authored by a current author ([17]) and the external ISTFT approach ([10]), are contextual or technical and do not carry the paper's central claim. Section 5.1 explicitly acknowledges that only reliability, not validity, is assessed and calls for future validity evaluation, which is a stated limitation rather than a circular move. The concern that low ISA may reflect inter-seed model variability rather than attribution unreliability is a substantive validity threat, but it is not an instance of the paper defining its conclusion into existence or importing a result from self-citation. Therefore no circularity is found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the assumption that inter-seed agreement measures method reliability, on hand-chosen metric settings (top-p=20%, silence baseline, 10ms perturbation spans, 20ms STFT hop), and on generalization from a single backbone and three datasets. No new entities are postulated.

free parameters (5)
  • top-p percentage in ISA = 20%
    ISA is computed as overlap of top 20% attribution indices (Eq. 1); the choice of 20% affects all reported ISA values and is not derived from any principle.
  • silence baseline = silence (zeros)
    Integrated Gradients, LIME, and Feature Ablation use silence as baseline input for waveform, spectrogram, and CNN embedding; baseline choice can materially alter attribution scores.
  • perturbation resolution = 10 ms spans for waveform perturbations
    Perturbation-based methods group waveform input into 10ms spans due to computational constraints; coarser resolution may affect LIME/FA reliability.
  • STFT/ISTFT hop length = 320 samples (20 ms)
    Spectrogram and CNN embedding attribution use hop length 320 to match wav2vec2's 20ms feature extractor; this determines frame boundaries for aggregation.
  • word mean-pooling = per-word mean
    Word-level aggregation mean-pools attribution scores per word and discards non-word segments; alternative pooling could change ISA.
assumptions (5)
  • domain assumption Randomly seeded fine-tuned runs of the same architecture and pretrained checkpoint use comparable internal features when their predictions agree.
    Section 3 states this assumption explicitly to justify ISA as a reliability measure; if false, attribution disagreement indicates model variance rather than method unreliability.
  • domain assumption Inter-seed agreement of top-20% attribution indices is a valid operationalization of reliability.
    The paper equates reliability with ISA (Eq. 1) and never validates ISA against an external standard; this is a definitional choice.
  • domain assumption Montreal Forced Aligner word boundaries are accurate enough for word-level aggregation and perturbation.
    Word-level conditions in Section 3.3 depend on forced-alignment timestamps; alignment errors would blur word-level signals.
  • domain assumption wav2vec2-base with frozen CNN and projection layers is representative of standard speech classification models.
    All experiments use one backbone and one fine-tuning protocol; the abstract's 'generally unreliable' claim generalizes beyond this setup.
  • domain assumption Captum's implementations of Saliency, Integrated Gradients, LIME, and Feature Ablation are correct and standard.
    Attribution scores are computed with Captum (Section 3.2); results depend on library defaults.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the reliability of feature attribution methods for speech classification." pith.science (2026). https://pith.science/paper/FLCRZSLN

@misc{pith2026250516406,
  author       = {Pith},
  title        = {Pith review of: On the reliability of feature attribution methods for speech classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLCRZSLN}},
  note         = {Machine review of arXiv:2505.16406}
}
read the original abstract

As the capabilities of large-scale pre-trained models evolve, understanding the determinants of their outputs becomes more important. Feature attribution aims to reveal which parts of the input elements contribute the most to model outputs. In speech processing, the unique characteristics of the input signal make the application of feature attribution methods challenging. We study how factors such as input type and aggregation and perturbation timespan impact the reliability of standard feature attribution methods, and how these factors interact with characteristics of each classification task. We find that standard approaches to feature attribution are generally unreliable when applied to the speech domain, with the exception of word-aligned perturbation methods when applied to word-based classification tasks.

Figures

Figures reproduced from arXiv: 2505.16406 by the authors.

Figure 1
Figure 1. Distributions of ISA scores without aggregation. The rows indicate different input feature types, the columns are dif￾ferent tasks. Within each panel, each boxplot shows results from different attribution methods and the y-axis is the ISA score. The red dotted line indicates the randomly shuffled baseline. IG: In￾tegrated Gradients, FA: Feature Ablation [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Distributions of ISA scores for the CNN embedding in￾put type, at different levels of aggregation. The rows are levels of granularity of aggregation, the columns are different tasks. Within each panel, each boxplot shows results from different at￾tribution methods and the y-axis is the ISA score. The red dot￾ted line indicates the randomly shuffled baseline. IG: Integrated Gradients, FA: Feature Ablation [PITH_FULL… view at source ↗
Figure 3
Figure 3. Distributions of ISA scores with perturbation operat￾ing directly on word-aligned segments. The rows indicate differ￾ent input feature types, the columns are different tasks. Within each panel, each boxplot report results from different attribu￾tion methods and the y-axis is the ISA score. The red dotted line indicates the randomly shuffled baseline. FA: Feature Ablation. ating directly on the word level. Here we se… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages

  1. [1]

    On the reliability of feature attribution methods for speech classification

    Introduction Large-scale self-supervised models such as wav2vec2 [1] and HuBERT [2] have shown impressive performance on various downstream speech processing tasks from automatic speech recognition to audio classification. As transformer models [3] have been increasingly adopted in speech processing, inter- pretability research for these models has also i...

    work page Pith review arXiv 2025

  2. [2]

    Related Work Feature attribution has evolved along with advances in machine learning, with methods originally developed to visualize salient features in computer vision [13, 14], and adapted further for natural language processing [15, 16, 17]. Unlike in speech pro- cessing, the limitations of feature attribution methods have been carefully studied in vis...

    work page

  3. [3]

    Methods For speech classification, pre-trained models are typically paired with a lightweight feedforward neural network as a classification head and fine-tuned using labeled data. Dur- ing fine-tuning, both the backbone model and classifier ad- just their weights to emphasize the most relevant input fea- tures and learned representations, maximizing clas...

    work page

  4. [4]

    Within each group, we present the effects of the varied conditions in applying feature attribution methods

    Results We organize the results into groups of comparisons. Within each group, we present the effects of the varied conditions in applying feature attribution methods. We visualize the central tendency (median) of the inter-seed agreement (ISA) scores as well as the spread around it via boxplots. An individual boxplot displays a set of 36 pairwise compari...

    work page

  5. [5]

    Discussion & Conclusion Our findings show that the naive application of standard fea- ture attribution methods to speech classification models gener- ally leads to poor reliability. When attributing to high-resolution input, regardless of specific input types such as waveform, spec- trogram or embeddings, even the most reliable of our methods, Integrated ...

    work page

  6. [6]

    Acknowledgements This publication is part of the projectInDeep: Interpreting Deep Learning Models for Text and Sound(with project number NW A.1292.19.399) of the National Research Agenda (NW A- ORC) program

    work page

  7. [7]

    Can We Trust Explainable AI Methods on ASR? An Eval- uation on Phoneme Recognition,

    ——, “Can We Trust Explainable AI Methods on ASR? An Eval- uation on Phoneme Recognition,” inICASSP 2024, Apr. 2024, pp. 10 296–10 300

    work page 2024

  8. [8]

    Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Represen- tations,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Represen- tations,” 2020

    work page 2020

Show all 39 references
  1. [9]

    HuBERT: Self-Supervised Speech Rep- resentation Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “HuBERT: Self-Supervised Speech Rep- resentation Learning by Masked Prediction of Hidden Units,” Jun. 2021

    2021

  2. [10]

    We use a hop length of 320 for the STFT and ISTFT transformations to keep the time resolution at 20ms to be consistent with the wav2vec2 model feature extrac- tor

    by prepending an inverse short-term Fourier transforma- tion (ISTFT) to the model. We use a hop length of 320 for the STFT and ISTFT transformations to keep the time resolution at 20ms to be consistent with the wav2vec2 model feature extrac- tor. For Integrated Gradients, LIME...

  3. [11]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” Dec. 2017

    2017

  4. [12]

    How Accents Confound: Probing for Accent Information in End-to-End Speech Recognition Systems,

    A. Prasad and P. Jyothi, “How Accents Confound: Probing for Accent Information in End-to-End Speech Recognition Systems,” inACL 2020, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, Eds. Online: ACL, Jul. 2020, pp. 3739–3753

    2020

  5. [13]

    AudioMNIST: Exploring Explainable Artificial Intelligence for audio analysis on a simple benchmark,

    S. Becker, J. Vielhaben, M. Ackermann, K.-R. M ¨uller, S. La- puschkin, and W. Samek, “AudioMNIST: Exploring Explainable Artificial Intelligence for audio analysis on a simple benchmark,” Journal of the Franklin Institute, vol. 361, no. 1, pp. 418–428, Jan. 2024

    2024

  6. [14]

    Explanations for Automatic Speech Recognition,

    X. Wu, P. Bell, and A. Rajan, “Explanations for Automatic Speech Recognition,” inICASSP 2023. Rhodes Island, Greece: IEEE, Jun. 2023, pp. 1–5

    2023

  7. [15]

    Phoneme Discretized Saliency Maps for Explainable Detection of AI- Generated V oice,

    S. Gupta, M. Ravanelli, P. Germain, and C. Subakan, “Phoneme Discretized Saliency Maps for Explainable Detection of AI- Generated V oice,” Sep. 2024

    2024

  8. [16]

    SPES: Spectrogram Perturbation for Explainable Speech-to-Text Generation,

    D. Fucci, M. Gaido, B. Savoldi, M. Negri, M. Cettolo, and L. Bentivogli, “SPES: Spectrogram Perturbation for Explainable Speech-to-Text Generation,” Nov. 2024

    2024

  9. [17]

    Investigating the Effectiveness of Explainability Methods in Parkinson’s Detection from Speech,

    E. Mancini, F. Paissan, P. Torroni, M. Ravanelli, and C. Sub- akan, “Investigating the Effectiveness of Explainability Methods in Parkinson’s Detection from Speech,” Nov. 2024

    2024

  10. [18]

    Explaining speech classification models via word-level au- dio segments and paralinguistic features,

    E. Pastor, A. Koudounas, G. Attanasio, D. Hovy, and E. Bar- alis, “Explaining speech classification models via word-level au- dio segments and paralinguistic features,” inEACL 2024, Y . Gra- ham and M. Purver, Eds., Mar. 2024

    2024

  11. [19]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. Mcleavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inProceedings of the 40th ICML. PMLR,

  12. [20]

    A song of (dis)agreement: Evaluating the evaluation of explainable artificial intelligence in natural language processing,

    M. Neely, S. F. Schouten, M. J. R. Bleeker, and A. Lucic, “A song of (dis)agreement: Evaluating the evaluation of explainable artificial intelligence in natural language processing,” inHHAI, 2022

    2022

  13. [21]

    Visualizing and Understanding Con- volutional Networks,

    M. D. Zeiler and R. Fergus, “Visualizing and Understanding Con- volutional Networks,” inComputer Vision – ECCV 2014. Cham: Springer International Publishing, 2014

    2014

  14. [22]

    Deep inside convolutional networks: Visualising image classification models and saliency maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: Visualising image classification models and saliency maps,” inICLR 2014, Workshop Track Proceedings, Y . Bengio and Y . LeCun, Eds., 2014. [Online]. Available: http://arxiv.org/abs/1312.6034

    2014 arXiv

  15. [23]

    Explaining by remov- ing: A unified framework for model explanation,

    I. Covert, S. M. Lundberg, and S.-I. Lee, “Explaining by remov- ing: A unified framework for model explanation,”J. Mach. Learn. Res., vol. 22, pp. 209:1–209:90, 2020

    2020

  16. [24]

    The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?

    J. Bastings and K. Filippova, “The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?” inBlackboxNLP 2020, A. Alishahi, Y . Belinkov, G. Chrupała, D. Hupkes, Y . Pinter, and H. Sajjad, Eds. Online: ACL, Nov. 2020, pp. 149–155

    2020

  17. [25]

    Exploring the role of BERT token representations to explain sentence probing results,

    H. Mohebbi, A. Modarressi, and M. T. Pilehvar, “Exploring the role of BERT token representations to explain sentence probing results,” inEMNLP 2021, M.-F. Moens, X. Huang, L. Specia, and S. W.-t. Yih, Eds. Online and Punta Cana, Dominican Republic: ACL, Nov. 2021, pp. 792–806

    2021

  18. [26]

    Evaluating explana- tions: How much do explanations from the teacher aid students?

    D. Pruthi, R. Bansal, B. Dhingra, L. Baldini Soares, M. Collins, Z. C. Lipton, G. Neubig, and W. W. Cohen, “Evaluating explana- tions: How much do explanations from the teacher aid students?” TACL, vol. 10, 2022

    2022

  19. [27]

    The disagreement problem in explainable machine learning: A practitioner’s perspective,

    S. Krishna, T. Han, A. Gu, S. Wu, S. Jabbari, and H. Lakkaraju, “The disagreement problem in explainable machine learning: A practitioner’s perspective,”TMLR, 2024

    2024

  20. [28]

    Common voice: A massively-multilingual speech corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the 12th LREC. ELRA, May 2020. [Online]. Available: https://aclanthology.org/2020.lrec-1.520/

    2020

  21. [29]

    When explanations lie: Why many modified bp attributions fail,

    L. Sixt, M. Granz, and T. Landgraf, “When explanations lie: Why many modified bp attributions fail,” inICML, 2019

    2019

  22. [30]

    Impossibil- ity theorems for feature attribution,

    B. Bilodeau, N. Jaques, P. W. Koh, and B. Kim, “Impossibil- ity theorems for feature attribution,”Proceedings of the National Academy of Sciences of the United States of America, vol. 121, 2022

    2022

  23. [31]

    “Will you find these shortcuts?

    J. Bastings, S. Ebert, P. Zablotskaia, A. Sandholm, and K. Filip- pova, ““Will you find these shortcuts?” a protocol for evaluating the faithfulness of input salience methods for text classification,” inEMNLP 2022, Y . Goldberg, Z. Kozareva, and Y . Zhang, Eds. ACL, Dec. 2022

    2022

  24. [32]

    Dy- namic Multi-granularity Attribution Network for Aspect-based Sentiment Analysis,

    Y . Chen, K. Zhang, F. Hu, X. Wang, R. Li, and Q. Liu, “Dy- namic Multi-granularity Attribution Network for Aspect-based Sentiment Analysis,” inEMNLP, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: ACL, 2024

    2024

  25. [33]

    Un- derstanding and Visualizing Raw Waveform-Based CNNs,

    H. Muckenhirn, V . Abrol, M. Magimai-Doss, and S. Marcel, “Un- derstanding and Visualizing Raw Waveform-Based CNNs,” inIn- terspeech 2019. ISCA, Sep. 2019

    2019

  26. [34]

    “Why Should I Trust You?

    M. T. Ribeiro, S. Singh, and C. Guestrin, ““Why Should I Trust You?”: Explaining the predictions of any classifier,” inProceed- ings of the 22nd ACM SIGKDD ICKDDM, 2016

    2016

  27. [35]

    TIMIT Acoustic-Phonetic Continuous Speech Cor- pus,

    Garofolo, John S., Lamel, Lori F., Fisher, William M., Pal- lett, David S., Dahlgren, Nancy L., Zue, Victor, and Fiscus, Jonathan G., “TIMIT Acoustic-Phonetic Continuous Speech Cor- pus,” 1993

    1993

  28. [37]

    Speech Model Pre-training for End-to-End Spoken Language Understanding,

    L. Lugosch, M. Ravanelli, P. Ignoto, V . S. Tomar, and Y . Bengio, “Speech Model Pre-training for End-to-End Spoken Language Understanding,” Jul. 2019

    2019

  29. [38]

    Captum: A unified and generic model interpretability library for PyTorch,

    N. Kokhlikyan, V . Miglani, M. Martin, E. Wang, B. Alsallakh, J. Reynolds, A. Melnikov, N. Kliushkina, C. Araya, S. Yan, and O. Reblitz-Richardson, “Captum: A unified and generic model interpretability library for PyTorch,” 2020

    2020

  30. [39]

    Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,” inInterspeech 2017. ISCA, Aug. 2017, pp. 498–502

    2017

  31. [2023]

    Available: https://proceedings.mlr.press/v202/ radford23a.html

    [Online]. Available: https://proceedings.mlr.press/v202/ radford23a.html

Pith tools

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