Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Echoes of Phonetics: Unveiling Relevant Acoustic Cues for ASR via Feature Attribution

T0 review · 3 major / 5 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 A modern Conformer-based ASR model relies on the same spectro-temporal cues that matter to human listeners: vowels' first two formants, sibilant spectra, and plosive bursts.

desk verdict A solid descriptive interpretability study of a Conformer ASR model, but the manually chosen 3% saliency threshold and missing statistical tests keep it at 'plausible' rather than 'established'. read the letter →

arxiv 2506.02181 v1 pith:HSNA5MCY submitted 2025-06-02 cs.CL cs.AI

Dennis Fucci , Marco Gaido , Matteo Negri , Mauro Cettolo , Luisa Bentivogli This is my paper
classification cs.CLcs.AI
keywords automaticspeechrecognitionfeatureattributionexplainabilitysaliencymapsacousticphoneticsConformerspectralmatchTIMIT
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 what a state-of-the-art, Conformer-based English ASR model actually listens to, by attributing each predicted word's importance to specific spectrogram elements and checking those saliency maps against phonetic knowledge. The authors find that the model relies on vowels over their full duration, weighting the first two formants (F1 and F2) more than higher formants; that it captures the spectral peaks of sibilant fricatives far better than those of non-sibilant fricatives; and that in plosives it prioritizes the release phase, especially burst characteristics. These alignments are strongest when the acoustic cue is spectrally well defined and weaker for diffuse cues, suggesting the model's listening strategy is closer to human phonetics than earlier architectures were thought to be.

What carries the argument

The load-bearing object is SPES (Spectrogram Perturbation for Explainable Speech-to-Text), a feature attribution method that clusters spectrogram elements by energy, perturbs those clusters to measure their impact on each predicted subword, and outputs a saliency map per token. The paper binarizes each map to the top 3% of salient elements and aligns them with TIMIT phonetic segmentation, then computes two metrics: time coverage (TC), the share of a phoneme's frames that contain at least one salient element, and spectral match (SM), the share of measured formants or spectral peaks that fall on salient elements. These two metrics carry the entire argument: all differences between vowels and consonants, sibilants and non-sibilants, and closure versus release are read off them.

What would settle it

On the same TIMIT SX subset, recompute time coverage and spectral match at saliency cutoffs of 1% and 10%; if the vowel-over-consonant and sibilant-over-non-sibilant gaps invert or vanish, the conclusions are an artifact of the 3% threshold rather than of model behavior.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a modern 113M-parameter Conformer encoder-decoder ASR model, trained on four large English corpora and tested on TIMIT, assigns prediction-relevant saliency in a way that tracks established acoustic-phonetic cues. For vowels, time coverage approaches 100% and spectral match is highest for F1 and F2 (averaging 71.4/64.8 and 73.7/72.4 for male/female speakers) and lower for F3 and F4, with systematically higher scores for male vowels. For fricatives, sibilants such as /s, z, S, Z/ show strong spectral-match scores and clear high-frequency saliency peaks, while non-sibilants such as /f, v, T, D/ show flat, low-density saliency; for plosives, saliency concentrates on the release phase rather than the closure, and burst spectral peaks are matched for velar and alveolar stops but poorly for labials. The authors conclude that the model's reliance on acoustic cues aligns with human speech perception where cues are well defined and degrades where they are not.

Load-bearing premise

The analysis keeps only the 3 percent of spectrogram elements deemed most important, and that cutoff was chosen by eye; the paper's comparisons all depend on that choice.

Editorial extensions

If this is right

  • If the central claim holds, ASR interpretability tools can move from neuron-level probing to phonetically meaningful statements: explaining a prediction means pointing to a formant, a frication noise band, or a burst.
  • The model's strong reliance on F1 and F2 predicts that perturbing or masking those formant regions will degrade recognition more than perturbing F3/F4 or consonant closures, which can be tested directly.
  • Because saliency for male vowels systematically exceeds that for female vowels, the paper predicts a link between feature-attribution patterns and known gender performance gaps in ASR.
  • Weak spectral match for non-sibilant fricatives and labial plosives implies those phoneme classes are the likeliest failure points under noise or channel conditions that blur their diffuse cues.

Reading between the lines

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

  • The top-3% threshold is the main knob: re-running the same pipeline at other cutoffs would reveal whether the reported phoneme-class contrasts are a property of the model or an artifact of binarization.
  • If the alignment-with-human-phonetics interpretation is right, then models trained with phonetic inductive biases (e.g., formant-based augmentation) should show even sharper saliency, while adversarially masking F1/F2 should expose the model's dependence.
  • The same attribution pipeline applied to a self-supervised speech encoder or a different decoder family would separate architectural effects from training-data effects: whether the Conformer, the CTC/decoder objective, or the English corpora produce the pattern.
  • The gender asymmetry could be probed by resynthesizing identical vowels with male vs female F0/formant values to see if the saliency shift is physiological rather than sociolinguistic.
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 / 5 minor

Summary. This paper applies the SPES feature-attribution method to a 113M-parameter Conformer-based ASR model trained on CommonVoice, LibriSpeech, TED-LIUM, and VoxPopuli, and analyzes saliency maps on error-free TIMIT SX sentences. It defines time coverage (TC) and spectral match (SM) to test whether the model's salient input regions align with phonetic cues: full vowel duration and F1/F2 formants, sibilant frication spectra, and plosive release bursts. The main claims are that vowels receive near-complete time coverage, sibilants are better matched spectrally than non-sibilants, plosive release is more salient than closure, and alignment is stronger for male than female vowels.

Significance. The study is a useful step beyond earlier analyses of outdated architectures, since it evaluates a modern Conformer/Transformer model on a wider phoneme inventory and provides code. The TC/SM metrics are simple and interpretable, and several qualitative patterns (e.g., low SM for /f/ and /v/, high SM for /S/) are visible in the tables. However, the headline time-coverage result is confounded by the arbitrary top-3% threshold and the lack of a chance baseline, and the gender and class differences are not supported by significance testing. If the threshold sensitivity and statistical issues are addressed, the conclusions about phonetic alignment would be genuinely informative for the interpretability community.

major comments (3)
  1. [Section 2, TC definition and Figure 2] The binary map S' retains the top 3% of elements based on 'manual inspection', and this single threshold drives the time-coverage results in Figure 2. Because F=80 (Section 3), a null map with the same number of salient bins placed uniformly at random would already produce an expected TC of 1-(0.97)^80 ≈ 91% per frame. The near-100% values for vowels are therefore close to chance, and the word-level reference line in Figure 2 is not a chance baseline. Please report TC and SM across a range of thresholds (e.g., 0.5%, 1%, 2%, 5%, 10%) and against a permutation null that preserves the total number of salient bins, and state whether the vowel/sibilant/release patterns are robust.
  2. [Section 4.2, Tables 1 and 2] The claims that the model captures sibilant spectra better than non-sibilants, that release bursts differ by place of articulation, and that vowel saliency is stronger for male speech rest on point estimates without confidence intervals or significance tests. For example, average F2 spectral match is 73.7 for men and 72.4 for women, a difference unlikely to be reliable without error bars; the gender claim in the abstract is directly based on such small differences. Add per-instance bootstrap or permutation confidence intervals (or at least effect sizes and counts) for the main TC and SM comparisons.
  3. [Section 3, Data paragraph] The analysis is restricted to the 2,191 error-free TIMIT SX predictions out of 3,150, while the abstract and conclusions are phrased about the model in general ('the ASR model relies...'). Please either limit the claims to correctly predicted utterances or analyze whether the pattern changes on the full set, including errors, which would also speak to the robustness gaps the introduction motivates.
minor comments (5)
  1. [Section 1] The introduction contains typos: 'ona wider range' and 'bya modern' should be 'on a wider range' and 'by a modern'.
  2. [Section 2] The phrase 'mean-standard normalization' is ambiguous; it should be described as z-score normalization, or the specific normalization formula should be given.
  3. [Tables 1 and 2] The tables do not report the number of phoneme instances per cell; uneven counts across phonemes and genders could affect the averages, so counts (or a statement that counts are balanced) should be provided.
  4. [Figure 2] The 'average coverage for a single word' reference line is not defined precisely; state how word-level TC was computed and whether it uses the same binary threshold.
  5. [Section 5, Limitations] The Limitations section does not mention the sensitivity of all quantitative results to the top-3% threshold or the restriction to error-free predictions; these are more consequential for the paper's claims than the omitted spectral moments.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; only a minor self-citation of the SPES attribution method, which is not load-bearing for the phonetic conclusions.

full rationale

The paper's central claims are descriptive measurements obtained by applying SPES, a previously published attribution method, to a trained Conformer ASR model. SPES is self-cited, but it is prior work used as a measurement instrument; its construction does not encode the phonetic conclusions about vowels' F1/F2, sibilant versus non-sibilant fricatives, or plosive release. The metrics TC and SM are defined directly from the thresholded saliency map, and no equation in the paper fits the target observations or equates a prediction with a fitted parameter. The top-3% saliency threshold is chosen by manual inspection and is not subjected to sensitivity analysis, which is a robustness limitation rather than a circular step, since the threshold is not fitted to the reported phonetic outcomes and the findings are not definitional consequences of the threshold choice. The self-citation of SPES is therefore minor and not load-bearing for the central claim, which has independent empirical content.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claims rest on a manually tuned saliency threshold and on domain assumptions about the attribution method, phoneme alignment, and acoustic cue extraction. No new entities are introduced.

free parameters (1)
  • top 3% saliency threshold = 3% of spectrogram elements per token
    Chosen by manual inspection of saliency distributions to 'best capture salient features'; all coverage and spectral match metrics depend on it (Section 2).
assumptions (3)
  • domain assumption SPES saliency maps faithfully indicate prediction-relevant spectrogram regions.
    The central conclusions interpret saliency as model reliance; SPES is assumed valid for this Conformer ASR model without a validation against alternative attribution methods here.
  • domain assumption Word-level max aggregation of token saliency maps preserves the per-phoneme relevance signal.
    Subword tokens cannot be aligned to TIMIT phonemes directly, so the authors take elementwise maxima across tokens, assuming no information loss (Section 3).
  • domain assumption Formant and spectral-peak measurements approximate the distinctive acoustic cues for each phoneme.
    Vowel formants from Parselmouth and mel-spectrogram peak bins are treated as the phonetic references for spectral match (Section 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Echoes of Phonetics: Unveiling Relevant Acoustic Cues for ASR via Feature Attribution." pith.science (2026). https://pith.science/paper/HSNA5MCY

@misc{pith2026250602181,
  author       = {Pith},
  title        = {Pith review of: Echoes of Phonetics: Unveiling Relevant Acoustic Cues for ASR via Feature Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HSNA5MCY}},
  note         = {Machine review of arXiv:2506.02181}
}
read the original abstract

Despite significant advances in ASR, the specific acoustic cues models rely on remain unclear. Prior studies have examined such cues on a limited set of phonemes and outdated models. In this work, we apply a feature attribution technique to identify the relevant acoustic cues for a modern Conformer-based ASR system. By analyzing plosives, fricatives, and vowels, we assess how feature attributions align with their acoustic properties in the time and frequency domains, also essential for human speech perception. Our findings show that the ASR model relies on vowels' full time spans, particularly their first two formants, with greater saliency in male speech. It also better captures the spectral characteristics of sibilant fricatives than non-sibilants and prioritizes the release phase in plosives, especially burst characteristics. These insights enhance the interpretability of ASR models and highlight areas for future research to uncover potential gaps in model robustness.

Figures

Figures reproduced from arXiv: 2506.02181 by the authors.

Figure 1
Figure 1. Mel-spectrogram (left), saliency map (center), and binary map (right) for the predicted token “us”. SPES clusters input spectrogram elements by energy, isolating features like harmonics and noise, and quantifies the impact of perturbing these clusters on yi. Finally, we convert Si into a binary map S ′ i ∈ {0, 1} T ×F , retaining only the most salient elements in Si (see [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of time coverage scores (TCϕ) for vowels (left), fricatives (center), and plosives (right). The dashed blue line represents the average coverage for a single word as a reference [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Distribution of the top 3% most salient elements across frequencies (DΦ) at the midpoint of /E/, with averaged formants (F1-F4, dotted lines), by gender. frequently located between 2 000 and 4 000 Hz for postalve￾olar fricatives (/S/, /Z/) and above 4 000 Hz for alveolar frica￾tives (/s/, /z/). These peaks are shifted toward higher frequen￾cies for women, as expected, with minimal differences between voiced (/z/, /Z… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Distribution of the top 3% most salient elements across frequencies (DΦ) at plosive release onset, by gender. and F2 than F3 and F4; ii) the model focuses on more defined frequency areas in sibilant fricatives than non-sibilant fricatives; iii) in plosives, release is …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Introduction State-of-the-art neural automatic speech recognition (ASR) sys- tems achieve impressive performance across languages but re- main opaque, offering limited insight into how they process acoustic information. To fill this gap, researchers have recently explored how speech models organize phonemic information, investigating their hidden states t...

    work page

  2. [2]

    Echoes of Phonetics: Unveiling Relevant Acoustic Cues for ASR via Feature Attribution

    Method Given an ASR model that takes a mel-spectrogramX∈R T×F as input—whereTis the number of time frames andFthe num- ber of frequency bins—and autoregressively predicts a token se- quencey= (y 0, y1, . . . , yI )of lengthIbased onXand prior tokens(y 0, . . . , yi−1), we investigate the relationship between Xand each predicted tokeny i. To this end, we u...

    work page Pith review arXiv 2025

  3. [3]

    Output text is encoded into BPE

    Experimental Setting ASR Model.We train an autoregressive encoder-decoder ASR model that processes log-compressed mel-filterbank features (80channels), computed over25ms windows with a10ms stride using PyKaldi [23]. Output text is encoded into BPE

    work page

  4. [4]

    Time Fig

    Results 4.1. Time Fig. 2 shows the distributions of thetime coverage(TC ϕ) scores. Distinct trends emerge between vowels and other phonemes. Forvowels, the TC ϕ scores are highly concentrated near100% with minimal variability. Their scores—except for outliers—exceed the average coverage for single words (used as a reference), showing that the model focuse...

    work page

  5. [5]

    The encoder layers use a convolution kernel size of31, an em- bedding size of512, and a linear layer hidden size of 2 048

    The processed features are then passed to a12-layer Con- former encoder [21] and a6-layer Transformer decoder [26]. The encoder layers use a convolution kernel size of31, an em- bedding size of512, and a linear layer hidden size of 2 048. The model comprises113M parameters and is implemented using the fairseq-S2T framework [27]. Training is conducted on C...

    work page

  6. [6]

    Acknowledgements The work presented in this paper has been funded from the PNRR project FAIR - Future AI Research (PE00000013), un- der the NRRP MUR program funded by the NextGenerationEU, and from the European Union’s Horizon research and innova- tion programme under grant agreement No 101135798, project Meetween (My Personal AI Mediator for Virtual MEET...

    work page

  7. [7]

    This is the first in-depth analysis of saliency maps in relation to fine-grained acoustic patterns across three phoneme classes

    Conclusions Using a feature attribution methodology, we analyzed the im- pact of distinctive acoustic cues of vowels, fricatives, and plo- sives in an English ASR model with a Conformer-based archi- tecture. This is the first in-depth analysis of saliency maps in relation to fine-grained acoustic patterns across three phoneme classes. We found that, in ge...

    work page

  8. [8]

    Understanding the representa- tion and computation of multilayer perceptrons: A case study in speech recognition,

    T. Nagamine and N. Mesgarani, “Understanding the representa- tion and computation of multilayer perceptrons: A case study in speech recognition,” inICML, 2017

    work page 2017

Show all 46 references
  1. [9]

    Analyzing phonetic and graphemic representations in end-to-end automatic speech recog- nition,

    Y . Belinkov, A. Ali, and J. Glass, “Analyzing phonetic and graphemic representations in end-to-end automatic speech recog- nition,” inInterspeech, 2019

    2019

  2. [10]

    Probing phoneme, language and speaker information in unsupervised speech representations,

    M. de Seyssel, M. Lavechin, Y . Adi, E. Dupoux, and G. Wis- niewski, “Probing phoneme, language and speaker information in unsupervised speech representations,” inInterspeech, 2022

    2022

  3. [11]

    Domain-Informed Probing of wav2vec 2.0 Embeddings for Pho- netic Features,

    P. Cormac English, J. D. Kelleher, and J. Carson-Berndsen, “Domain-Informed Probing of wav2vec 2.0 Embeddings for Pho- netic Features,” inSIGMORPHON Workshop, 2022

    2022

  4. [12]

    Probing self- supervised speech models for phonetic and phonemic informa- tion: A case study in aspiration,

    K. Martin, J. Gauthier, C. Breiss, and R. Levy, “Probing self- supervised speech models for phonetic and phonemic informa- tion: A case study in aspiration,” inInterspeech, 2023

    2023

  5. [13]

    Self-supervised speech representations are more phonetic than semantic,

    K. Choi, A. Pasad, T. Nakamura, S. Fukayama, K. Livescu, and S. Watanabe, “Self-supervised speech representations are more phonetic than semantic,” inInterspeech, 2024

    2024

  6. [14]

    Encoding of phonol- ogy in a recurrent neural model of grounded speech,

    A. Alishahi, M. Barking, and G. Chrupała, “Encoding of phonol- ogy in a recurrent neural model of grounded speech,” inCoNLL, 2017

    2017

  7. [15]

    Learn- ing weakly supervised multimodal phoneme embeddings,

    R. Chaabouni, E. Dunbar, N. Zeghidour, and E. Dupoux, “Learn- ing weakly supervised multimodal phoneme embeddings,” inIn- terspeech, 2017

    2017

  8. [16]

    Acous- tic characteristics of American English vowels,

    J. Hillenbrand, L. A. Getty, M. J. Clark, and K. Wheeler, “Acous- tic characteristics of American English vowels,”The Journal of the Acoustical Society of America, vol. 97, no. 5, pp. 3099–3111, 1995

    1995

  9. [17]

    Neuron Activation Profiles for Interpreting Convolutional Speech Recognition Models,

    A. Krug, R. Knaebel, and S. Stober, “Neuron Activation Profiles for Interpreting Convolutional Speech Recognition Models,” in NeurIPS Workshop IRASL, 2018

    2018

  10. [18]

    Interpretable Convolutional Filters with SincNet,

    M. Ravanelli and Y . Bengio, “Interpretable Convolutional Filters with SincNet,” inNeurIPS Workshop IRASL, 2018

    2018

  11. [19]

    Introspection for convolutional automatic speech recognition,

    A. Krug and S. Stober, “Introspection for convolutional automatic speech recognition,” inEMNLP Workshop BlackboxNLP, 2018

    2018

  12. [20]

    End-to-end acoustic modeling using convolutional neural networks for HMM- based automatic speech recognition,

    D. Palaz, M. Magimai-Doss, and R. Collobert, “End-to-end acoustic modeling using convolutional neural networks for HMM- based automatic speech recognition,”Speech Communication, vol. 108, pp. 15–32, 2019

    2019

  13. [21]

    Gradient-Adjusted Neuron Activa- tion Profiles for Comprehensive Introspection of Convolutional Speech Recognition Models,

    A. Krug and S. Stober, “Gradient-Adjusted Neuron Activa- tion Profiles for Comprehensive Introspection of Convolutional Speech Recognition Models,”arXiv preprint arXiv:2002.08125, 2020

    2002 arXiv

  14. [22]

    Directly Comparing the Listening Strategies of Humans and Machines,

    V . A. Trinh and M. Mandel, “Directly Comparing the Listening Strategies of Humans and Machines,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 312–323, 2021

    2021

  15. [23]

    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,”arXiv preprint arXiv:2411.01710, 2024

    2024 arXiv

  16. [24]

    with a vocabulary size of 8 000, using SentencePiece [25]. The input features are normalized via utterance-level Cepstral Mean and Variance Normalization and downsampled by a fac- tor of4through two1D convolutional layers with a stride of 2Formant measurements are carried out ...

  17. [25]

    Burst and Transition Cues to V oicing Perception for Spo- ken Initial Stops by Impaired- and Normal-Hearing Listeners,

    S. Revoile, J. M. Pickett, L. D. Holden-Pitt, D. Talkin, and F. D. Brandt, “Burst and Transition Cues to V oicing Perception for Spo- ken Initial Stops by Impaired- and Normal-Hearing Listeners,” Journal of Speech, Language, and Hearing Research, vol. 30, no. 1, pp. 3–12, 1987

    1987

  18. [26]

    Acoustic cues of voiced and voiceless plosives for determining place of articulation,

    P. J. Jackson, “Acoustic cues of voiced and voiceless plosives for determining place of articulation,” inWorkshop on Consistent and Reliable Acoustic Cues for sound analysis, 2001

    2001

  19. [27]

    Acoustic characteristics of English fricatives,

    A. Jongman, R. Wayland, and S. Wong, “Acoustic characteristics of English fricatives,”The Journal of the Acoustical Society of America, vol. 108, no. 3, pp. 1252–1263, 2000

    2000

  20. [28]

    Acoustic characteristics of clearly spoken English fricatives,

    K. Maniwa, A. Jongman, and T. Wade, “Acoustic characteristics of clearly spoken English fricatives,”The Journal of the Acousti- cal Society of America, vol. 125, no. 6, 2009

    2009

  21. [29]

    Conformer: Convolution-augmented Transformer for Speech Recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented Transformer for Speech Recognition,” in Interspeech, 2020

    2020

  22. [30]

    Introducing Parsel- mouth: A Python interface to Praat,

    Y . Jadoul, B. Thompson, and B. de Boer, “Introducing Parsel- mouth: A Python interface to Praat,”Journal of Phonetics, vol. 71, pp. 1–15, 2018

    2018

  23. [31]

    Pykaldi: A Python Wrapper for Kaldi,

    D. Can, V . R. Martinez, P. Papadopoulos, and S. S. Narayanan, “Pykaldi: A Python Wrapper for Kaldi,” inICASSP, 2018

    2018

  24. [32]

    Neural Machine Transla- tion of Rare Words with Subword Units,

    R. Sennrich, B. Haddow, and A. Birch, “Neural Machine Transla- tion of Rare Words with Subword Units,” inACL, 2016

    2016

  25. [33]

    SentencePiece: A simple and lan- guage independent subword tokenizer and detokenizer for Neural Text Processing,

    T. Kudo and J. Richardson, “SentencePiece: A simple and lan- guage independent subword tokenizer and detokenizer for Neural Text Processing,” inEMNLP, 2018

    2018

  26. [34]

    Attention is All you Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is All you Need,” inNeurIPS, 2017

    2017

  27. [35]

    Fairseq S2T: Fast Speech-to-Text Modeling with Fairseq,

    C. Wang, Y . Tang, X. Ma, A. Wu, D. Okhonko, and J. Pino, “Fairseq S2T: Fast Speech-to-Text Modeling with Fairseq,” in AACL-IJCNLP, 2020

    2020

  28. [36]

    Com- mon V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Hen- retty, R. Morais, L. Saunders, F. Tyers, and G. Weber, “Com- mon V oice: A Massively-Multilingual Speech Corpus,” inLREC, 2020

    2020

  29. [37]

    Lib- rispeech: An ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An ASR corpus based on public domain audio books,” inICASSP, 2015

    2015

  30. [38]

    TED-LIUM 3: Twice as Much Data and Corpus Repartition for Experiments on Speaker Adaptation,

    F. Hernandez, V . Nguyen, S. Ghannay, N. Tomashenko, and Y . Est`eve, “TED-LIUM 3: Twice as Much Data and Corpus Repartition for Experiments on Speaker Adaptation,” inSpeech and Computer, 2018

    2018

  31. [39]

    V oxPopuli: A Large- Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation,

    C. Wang, M. Riviere, A. Lee, A. Wu, C. Talnikar, D. Haziza, M. Williamson, J. Pino, and E. Dupoux, “V oxPopuli: A Large- Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation,” inACL-IJCNLP, 2021

    2021

  32. [40]

    Rethinking the Inception Architecture for Computer Vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,” in CVPR, 2016

    2016

  33. [41]

    Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” inICML, 2006

    2006

  34. [42]

    Adam: A Method for Stochastic Opti- mization,

    D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Opti- mization,” inICLR, 2015

    2015

  35. [43]

    SpecAugment: A Simple Data Aug- mentation Method for Automatic Speech Recognition,

    D. S. Park, W. Chanet al., “SpecAugment: A Simple Data Aug- mentation Method for Automatic Speech Recognition,” inInter- speech, 2019

    2019

  36. [44]

    Darpa timit acoustic-phonetic continuous speech corpus cd-rom TIMIT,

    J. Garofolo, L. Lamel, W. Fisher, J. Fiscus, D. Pallett, and N. Dahlgren, “Darpa timit acoustic-phonetic continuous speech corpus cd-rom TIMIT,” 1993-02-01 1993

    1993

  37. [45]

    Twists, humps, and pebbles: Multilingual speech recognition models exhibit gen- der performance gaps,

    G. Attanasio, B. Savoldi, D. Fucci, and D. Hovy, “Twists, humps, and pebbles: Multilingual speech recognition models exhibit gen- der performance gaps,” inEMNLP, 2024

    2024

  38. [46]

    Burst spectrum as a cue for the stop voicing contrast in American English,

    E. Chodroff and C. Wilson, “Burst spectrum as a cue for the stop voicing contrast in American English,”The Journal of the Acous- tical Society of America, vol. 136, no. 5, pp. 2762–2772, 2014

    2014

Pith tools

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