Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Automatic Speech Recognition of African American English: Lexical and Contextual Effects

T0 review · 4 major / 6 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 Two African American English sound changes measurably raise ASR error rates.

desk verdict Useful feature-level WER analysis of AAE variables; H1 solid, H2 undercut by LM trained on the test corpus. read the letter →

arxiv 2506.06888 v2 pith:AUAFGSBM submitted 2025-06-07 cs.CL cs.SDeess.AS

Hamid Mojarad , Kevin Tang This is my paper
classification cs.CLcs.SDeess.AS
keywords automaticspeechrecognitionAfricanAmericanEnglishconsonantclusterreductionING-reductionlexicalneighborhoodeffectcontextualpredictabilitywav2vec2.0languagemodel
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 two common phonological features of African American English, consonant cluster reduction (dropping a final stop from a cluster, as in cold pronounced col) and ING-reduction (pronouncing -ing with an alveolar n rather than a velar ng), are measurable causes of automatic speech recognition errors. Using the CORAAL Washington, DC corpus transcribed by wav2vec 2.0 with and without an external language model, the authors find a small but statistically significant increase in word error rate for reduced pronunciations, with regression estimates between 0.021 and 0.040. They also find that the system without a language model attributes 7.9% of target-word errors to lexical neighbor confusions, while adding a language model cuts that share to 3.3%, supporting the claim that language models shift errors from neighborhood confusions toward other sources. If correct, the result gives concrete phonological targets for reducing dialect bias in speech recognition and quantifies how much of the bias is linguistic rather than acoustic.

What carries the argument

The argument rides on three linked tools. First, forced alignment with pronunciation expansion: the Montreal Forced Aligner is trained on the full DCA audio set and given both original and reduced dictionary pronunciations for CCR- and ING-prone words, so every target token is classified as original or reduced rather than by hand. Second, wav2vec 2.0 large-960h runs twice, once as a pure end-to-end system and once with a 5-gram KenLM language model trained on the Washington, DC CORAAL subcorpora, allowing the same audio to be scored with and without contextual prediction. Third, lexical neighborhood status is defined by Luce's criterion, one phoneme substitution, deletion, or addition, computed by Levenshtein distance between the target's pronunciation and the ASR output, so a misrecognition counts as a neighbor error exactly when it lands on a phonological competitor of the target.

What would settle it

Train the same 5-gram KenLM on only the DCB subcorpus and any held-out DCA recordings not used in scoring, then re-run the with-LM transcription. If the neighbor-error share stays near 3.3%, the contextual-predictability claim is supported; if it returns toward 7.9%, the original drop was driven by the LM having seen the test transcripts.

Watch

Extended reading notes

Core claim

The paper's central claim is that the presence of CCR and ING-reduction in AAE speech increases ASR misrecognition, and that end-to-end ASR without an external language model is more influenced by lexical neighborhood density and less by contextual predictability than the same system with a language model. The evidence comes from 34 hours of CORAAL DCA interviews, with target words aligned by the Montreal Forced Aligner using expanded pronunciations so each token is tagged as original or reduced. Across overall, CCR-only, and ING-only datasets, the MFA Status variable (original vs reduced) has a significant positive effect on WER both with and without the LM, with estimates from 0.021 to 0.040. For the second hypothesis, neighbor errors, where the ASR output is one phoneme edit away from the target, fall from 1,006 of 12,734 incorrect transcriptions (7.9%) without the LM to 277 of 8,283 (3.3%) with the LM; the logistic mixed-effects model confirms a significant effect of ASR type across all datasets. The authors read this as evidence that an external LM supplies contextual predictability that resolves some acoustically ambiguous reduced forms.

Load-bearing premise

The load-bearing premise for the second hypothesis is that the 5-gram language model is genuinely external to the test set; in fact it is trained on the entire Washington, DC CORAAL data, including the DCA subcorpus that is later transcribed and scored, so the drop in neighbor errors could come from in-domain text exposure rather than general contextual predictability.

Editorial extensions

If this is right

  • Reduced CCR and ING pronunciations will keep inflating WER even when an external LM is added, so acoustic-level adaptation to these two variables is a concrete lever for reducing AAE ASR bias.
  • Adding an LM changes the error distribution: neighbor confusions drop from 7.9% to 3.3% of errors, meaning the remaining error burden is dominated by non-neighbor failures such as acoustic mismatch and other dialect features.
  • The age pattern among AAE speakers, highest WER for the under-19 group, reverses the usual child and elderly ASR pattern and suggests generational linguistic differences matter more than vocal-tract age in this corpus.
  • Gender has no significant effect on CCR and ING-prone word errors, so gender-based bias is not a substantial driver for these particular AAE variables.
  • Annotating phonological variation in training data, for example marking reduced tokens, is a directly suggested route to improving accuracy.

Reading between the lines

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

  • Editorial extension: the paper's LM is trained on the entire Washington, DC CORAAL data, which includes the same DCA subcorpus later transcribed and scored; the drop from 7.9% to 3.3% may therefore reflect in-domain text exposure rather than general contextual predictability.
  • Editorial extension: because reduced forms are shorter, they have denser lexical neighborhoods by construction; WER differences between original and reduced tokens could partly be an artifact of neighborhood density rather than of the phonological reduction itself.
  • Editorial extension: a practical testable extension is to fine-tune wav2vec 2.0 on CORAAL tokens tagged original versus reduced, then check whether WER for reduced forms drops below the LM-only baseline; this would determine whether the bias is learnable from acoustic data.
  • Editorial extension: the finding suggests that bias metrics comparing WER across dialects should control for lexical neighborhood density, otherwise dialect differences and word-length differences are conflated.
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 / 6 minor

Summary. This paper investigates whether two AAE phonological variables—consonant cluster reduction (CCR) and ING-reduction—increase ASR word error rate, and whether adding a 5-gram KenLM reduces lexical-neighborhood errors. Using the CORAAL DCA subcorpus, the authors force-align with MFA to label target words as original or reduced, transcribe with wav2vec 2.0 with and without an LM, and fit linear/logistic mixed-effects models. H1 finds a small but significant positive effect of reduced pronunciation on WER across datasets and LM conditions. H2 finds that neighbor errors drop from 7.9% to 3.3% when the LM is used.

Significance. If the results hold, the paper makes a useful contribution by quantifying two specific AAE phonological variables as error sources in a modern end-to-end ASR model and by showing that LM integration changes the error profile. Strengths include the use of a large naturalistic corpus, open data/code on OSF, and mixed-effects modeling that accounts for speaker and word variability. However, the support for H2 is weakened by the LM's training overlap with the test corpus, and the H1 predictor (MFA labels) is not validated against human coding. The empirical claims in the abstract are therefore only partially supported.

major comments (4)
  1. [Section 2.3 vs. Section 2.1] The 'external' 5-gram KenLM is trained on CORAAL's DCA and DCB subcorpora (Section 2.3), while the test set consists of DCA interviewee utterances (Sections 2.1 and 2.4). The LM therefore saw the test transcripts, including their exact 5-gram sequences. The drop in neighbor errors from 7.9% (1,006/12,734) to 3.3% (277/8,283) reported in Section 3.2.3 could be driven by in-domain text exposure rather than by general contextual predictability. The abstract's comparative claim about lexical-neighborhood versus contextual-predictability influence depends on the LM being external to the scored data and is not supported as stated. Please re-run with a properly held-out LM (e.g., trained only on DCB or on another corpus) or reframe the result as an in-domain LM effect.
  2. [Section 3.2.1 and Section 3.2.3] The contrast coding in Section 3.2.1 is without LM = -0.5 and with LM = 0.5, but Section 3.2.3 reports 'a consistent, significant positive effect' while listing negative coefficients (-2.1879 for ING, -1.2875 for overall, -0.8954 for CCR). A negative coefficient means the with-LM condition reduces neighbor errors, so the prose should say 'negative effect' or the coding should be reversed. In addition, because the LM is in-domain, these coefficients should not be interpreted as evidence for general contextual predictability.
  3. [Section 3.1.1] The dependent variable in H1 is called WER, but the observation-level definition is never stated. If WER is a binary word-level correct/incorrect outcome, a linear mixed-effects model is a linear probability model, which is questionable for binary data; if WER is an utterance-level rate, the model should account for the number of words per utterance (e.g., via weights or a binomial family). The reported estimates (0.021 to 0.040) are described as effect sizes, but their units are unclear without this definition. Please clarify how WER is computed and justify the model family.
  4. [Section 2.2 and Section 4] The MFA-based detection of CCR and ING-reduction is a load-bearing predictor for H1, but the paper acknowledges in Section 4 that it was not evaluated against human coding. Kendall et al. [6] validated ING coding, but not with the custom acoustic model trained here, and no validation is provided for CCR. Because the acoustic model was trained on the same DCA audio used for alignment, the reliability of the 'reduced' labels is uncertain. Please report a validation subset or a sensitivity analysis showing how misclassification would affect the H1 estimates.
minor comments (6)
  1. [Abstract] The phrase 'external Language Model' is misleading given that the LM is trained on the same Washington DC corpus used for scoring; please qualify this as an in-domain or same-corpus LM.
  2. [Section 2.2] The selection criteria for 'target words' are not specified; please state how CCR-prone and ING-prone words were identified and report the number of word types and tokens per feature.
  3. [Section 2.3] The paper says audio was segmented into 'chunks of at least 30 seconds,' but wav2vec 2.0 typically has a maximum input length; please clarify the chunking procedure and whether the full 34 hours of DCA audio were transcribed.
  4. [Section 3.2.3] Please report p-values as p < 0.001 and include confidence intervals for the odds ratios rather than only the raw coefficients, and correct the informal 'ps < 0.001' notation.
  5. [Section 4] The sentence 'This reduction can be attributed to the LM's ability to provide contextual predictability' overstates the finding; given the training-data overlap, a more cautious phrasing such as 'is consistent with' would be appropriate.
  6. [Figure 1] The caption does not describe the plot type or the meaning of boxes/whiskers; please add a description and define the WER calculation used in the figure.

Circularity Check

1 steps flagged · score 6.0 of 10

H2's 'external LM' is trained on the same DCA subcorpus used as the test set, contaminating the contextual-predictability comparison.

  1. fitted input called prediction [Section 2.3 (ASR transcription) with Section 2.1 and 2.4, evaluated in Section 3.2.3]
    "For this research, we specifically utilized the DCA (Washington, DC) subcorpus ... This version, trained on 960 hours of speech, was subsequently enhanced with an external 5-gram LM trained on CORAAL's DCA and DCB subcorpora (entire Washington DC data) using KenLM. ... To ensure a focused analysis of AAE features, we exclusively processed utterances from interviewees, excluding those of interviewers from the DCA dataset."

    The 5-gram KenLM is trained on the entire Washington DC CORAAL data (DCA plus DCB), while the H2 analysis is restricted to DCA interviewee utterances. Thus the LM's training text includes the very transcripts later used as the test ground truth. The reported drop in neighborhood errors from 7.9% (1,006/12,734) to 3.3% (277/8,283) when the LM is added is therefore not a clean measure of general contextual predictability: a 5-gram LM can memorize in-domain test sentences and assign them artificially high probability. Labelling the LM 'external' is only true relative to the acoustic model, not relative to the test corpus.

full rationale

The H1 result (CCR and ING-reduction increase WER) is self-contained: it compares MFA-detected reduced versus original pronunciations within the same ASR systems, with no reliance on an external fitted quantity for the core contrast. No load-bearing self-citation chain, ansatz smuggling, or renaming of known results appears; citations to the authors' prior work are contextual and not the basis of the main derivation. The significant circularity is confined to H2. Section 2.3 states that the 'external' LM was trained on CORAAL's DCA and DCB subcorpora, while Section 2.4 limits the test utterances to the DCA dataset. The LM has therefore seen the test transcripts, so the 7.9% to 3.3% reduction in neighborhood errors may reflect in-domain memorization rather than general contextual predictability. The paper's abstract and discussion attribute the reduction to the LM's ability to use context, but the experimental design cannot separate contextual generalization from test-corpus exposure. This is a partial, construction-level contamination of the paper's second central claim; H1 remains independently supported, so the overall score is 6 rather than higher.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the reliability of forced-alignment labels and on the validity of the LM comparison; the latter is weakened by training on the test corpus. Regression coefficients are estimated outputs, not fitted constants, so no free parameters are introduced.

assumptions (4)
  • domain assumption MFA with pronunciation expansion accurately labels CCR and ING-reduction in CORAAL.
    Used as ground truth for reduced vs original; authors acknowledge no human coding check for CCR.
  • ad hoc to paper The KenLM 5-gram trained on CORAAL DCA and DCB is an external LM for scoring DCA.
    Section 2.3: training data includes the test subcorpus, so 'external' is inaccurate.
  • domain assumption CORAAL DCA time-aligned orthographic transcriptions are accurate ground truth.
    Transcripts are the reference for WER; errors in them propagate to all measurements.
  • domain assumption wav2vec2-large-960h without an LM is representative of end-to-end ASR systems.
    Section 2.3: conclusions about end-to-end systems rest on this single model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatic Speech Recognition of African American English: Lexical and Contextual Effects." pith.science (2026). https://pith.science/paper/AUAFGSBM

@misc{pith2026250606888,
  author       = {Pith},
  title        = {Pith review of: Automatic Speech Recognition of African American English: Lexical and Contextual Effects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AUAFGSBM}},
  note         = {Machine review of arXiv:2506.06888}
}
read the original abstract

Automatic Speech Recognition (ASR) models often struggle with the phonetic, phonological, and morphosyntactic features found in African American English (AAE). This study focuses on two key AAE variables: Consonant Cluster Reduction (CCR) and ING-reduction. It examines whether the presence of CCR and ING-reduction increases ASR misrecognition. Subsequently, it investigates whether end-to-end ASR systems without an external Language Model (LM) are more influenced by lexical neighborhood effect and less by contextual predictability compared to systems with an LM. The Corpus of Regional African American Language (CORAAL) was transcribed using wav2vec 2.0 with and without an LM. CCR and ING-reduction were detected using the Montreal Forced Aligner (MFA) with pronunciation expansion. The analysis reveals a small but significant effect of CCR and ING on Word Error Rate (WER) and indicates a stronger presence of lexical neighborhood effect in ASR systems without LMs.

Figures

Figures reproduced from arXiv: 2506.06888 by the authors.

Figure 1
Figure 1. WER by MFA Status for CCR and ING Target Words (reference level: Neighbor Error), and contrast coding was ap￾plied to ASR Type (without LM: -0.5, with LM: 0.5). Mixed￾effects logistic regression was employed with Target Word and Speaker Id as random effects. Furthermore, for both random ef￾fects, we included random slopes for ASR Type to allow for the impact of ASR type on Neighborhood Status to vary for differ￾ent … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 35 canonical work pages

  1. [6]

    Acknowledgements 5 The work is part of HM’s PhD. Conceptualization, Methodol- ogy, Formal Analysis, Writing – original draft/review/editing: HM, KT; Data Curation, Investigation, Funding Acquisition, Project Administration, Software, Validation: HM; Resources, Supervision: KT

    work page

  2. [1]

    Automatic Speech Recognition of African American English: Lexical and Contextual Effects

    Introduction Addressing racial bias in ASR has recently become a significant area of concern. Given AAE as a minority dialect, this issue was phonetically confirmed by Koenecke et al. [1], who found that the average WER for white American speakers was sig- nificantly lower than that for AAE speakers across five promi- nent ASR systems. Morphosyntactic dis...

    work page Pith review arXiv 2025

  3. [2]

    accept” was represented as both “AH0 K S EH1 P T

    Methodology 2.1. Corpus The Corpus of Regional African American Language (CORAAL) [17] serves as the foundational dataset for this study, offering a comprehensive documentation of regional African American Language (AAL) varieties. The corpus provides rich linguistic resources, including audio recordings with time-aligned orthographic transcriptions in Te...

    work page

  4. [3]

    Original Pronunciation

    Analyses 3.1. H1: phonological reduction increases ASR errors 3.1.1. Variables In H1, WER was analyzed as the dependent variable, with MFA Status, AgeGroup, and Gender serving as fixed effect vari- ables. MFA Status served as a binary factor indicating whether the pronunciation was detected as original or reduced by MFA. To address potential non-independe...

    work page

  5. [4]

    The consistent positive effect of MFA Status across datasets indicates that AAE features signif- icantly influence ASR misrecognition

    Discussion Our study reveals notable insights into the performance of ASR systems when confronted with CCR and ING-reduction, as common AAE variations. The consistent positive effect of MFA Status across datasets indicates that AAE features signif- icantly influence ASR misrecognition. This effect still remains significant even when we recruit an external...

    work page 1968

  6. [5]

    Conclusion This study examined the performance of ASR systems, focus- ing on CCR and ING-reduction, two common phonological variations in AAE. Our findings underscore the persistent chal- lenges ASR systems face when transcribing dialectal speech, even with advanced architectures like wav2vec 2.0 and the in- tegration of LMs. First, our results confirmed ...

    work page

  7. [7]

    Racial disparities in automated speech recognition,

    A. Koenecke, A. Nam, E. Lake, J. Nudell, M. Quartey, Z. Menge- sha, C. Toups, J. R. Rickford, D. Jurafsky, and S. Goel, “Racial disparities in automated speech recognition,” Proceedings of the National Academy of Sciences , vol. 117, no. 14, pp. 7684–7689, 2020

    work page 2020

  8. [8]

    Understanding racial disparities in au- tomatic speech recognition: The case of habitual “be

    J. L. Martin and K. Tang, “Understanding racial disparities in au- tomatic speech recognition: The case of habitual “be”,” in Inter- speech 2020, 2020, pp. 626–630

    work page 2020

Show all 37 references
  1. [9]

    Uneven suc- cess: automatic speech recognition and ethnicity-related dialects,

    A. B. Wassink, C. Gansen, and I. Bartholomew, “Uneven suc- cess: automatic speech recognition and ethnicity-related dialects,” Speech Communication, vol. 140, pp. 50–70, 2022

    2022

  2. [10]

    Segmental phonology of African American English,

    E. R. Thomas and G. Bailey, “Segmental phonology of African American English,” inThe Oxford Handbook of African American Language. Oxford University Press, 07 2015

    2015

  3. [11]

    A comparative analysis of consonant clusters in En- glish and in Slovak,

    R. Gregov ´a, “A comparative analysis of consonant clusters in En- glish and in Slovak,” Bulletin of the Transilvania University of Brasov. Series IV: Philology and Cultural Studies , pp. 79–84, 2010

    2010

  4. [12]

    Considering performance in the au- tomated and manual coding of sociolinguistic variables: Lessons from variable (ING),

    T. Kendall, C. Vaughn, C. Farrington, K. Gunter, J. McLean, C. Tacata, and S. Arnson, “Considering performance in the au- tomated and manual coding of sociolinguistic variables: Lessons from variable (ING),” Frontiers in Artificial Intelligence , vol. 4, 2021

    2021

  5. [13]

    Recognizing spoken words: The neighborhood activation model,

    P. A. Luce and D. B. Pisoni, “Recognizing spoken words: The neighborhood activation model,” Ear and Hearing, vol. 19, no. 1, pp. 1–36, 1998

    1998

  6. [14]

    Revisiting word neighborhoods for speech recognition,

    P. Jyothi and K. Livescu, “Revisiting word neighborhoods for speech recognition,” in Proceedings of the 2014 Joint Meeting of SIGMORPHON and SIGFSM, ¨O. C ¸ etino˘glu, J. Heinz, A. Maletti, and J. Riggle, Eds. Baltimore, Maryland: Association for Com- putational Linguistics, J...

    2014

  7. [15]

    Which words are hard to recognize? prosodic, lexical, and disfluency factors that increase speech recognition error rates,

    S. Goldwater, D. Jurafsky, and C. D. Manning, “Which words are hard to recognize? prosodic, lexical, and disfluency factors that increase speech recognition error rates,” Speech Communication, vol. 52, no. 3, pp. 181–200, 2010

    2010

  8. [16]

    Clearpond: Cross-linguistic easy-access resource for phonological and ortho- graphic neighborhood densities,

    V . Marian, J. Bartolotti, S. Chabal, and A. Shook, “Clearpond: Cross-linguistic easy-access resource for phonological and ortho- graphic neighborhood densities,” PLOS ONE, vol. 7, no. 8, pp. 1–11, 08 2012

    2012

  9. [17]

    Contextualized end-to-end speech recognition with contextual phrase prediction network,

    K. Huang, A. Zhang, Z. Yang, P. Guo, B. Mu, T. Xu, and L. Xie, “Contextualized end-to-end speech recognition with contextual phrase prediction network,” in Interspeech 2023, 2023, pp. 4933– 4937

    2023

  10. [18]

    Improving speech recogni- tion with prompt-based contextualized ASR and LLM-based re- predictor,

    N. Manh Tien Anh and T. Ho Sy, “Improving speech recogni- tion with prompt-based contextualized ASR and LLM-based re- predictor,” in Interspeech 2024, 2024, pp. 737–741

    2024

  11. [19]

    Improving contextual recognition of rare words with an alternate spelling prediction model,

    J. Fox and N. Delworth, “Improving contextual recognition of rare words with an alternate spelling prediction model,” inInterspeech 2022, 2022, pp. 3914–3918

    2022

  12. [20]

    Lost in the middle: How language mod- els use long contexts,

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang, “Lost in the middle: How language mod- els use long contexts,” Transactions of the Association for Com- putational Linguistics, vol. 12, pp. 157–173, 2024

    2024

  13. [21]

    Assessing the importance of several acoustic properties to the perception of spontaneous speech,

    R. G. Podlubny, T. M. Nearey, G. Kondrak, and B. V . Tucker, “Assessing the importance of several acoustic properties to the perception of spontaneous speech,” The Journal of the Acoustical Society of America, vol. 143, no. 4, pp. 2255–2268, 04 2018. 5https://credit.niso.org/

    2018

  14. [22]

    Contextualized speech recognition: Rethinking second-pass rescoring with generative large language models,

    Y . Tang and A. K. H. Tung, “Contextualized speech recognition: Rethinking second-pass rescoring with generative large language models,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24 , K. Larson, Ed. International Joint ...

    2024

  15. [23]

    The Corpus of Regional African American Language,

    T. Kendall and C. Farrington, “The Corpus of Regional African American Language,” 2023, publisher: The Online Resources for African American Language Project. [Online]. Available: https://oraal.uoregon.edu/coraal

    2023

  16. [24]

    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,” in Proc. Interspeech 2017 , 2017, pp. 498–502

    2017

  17. [25]

    wav2vec 2.0: a framework for self-supervised learning of speech representa- tions,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: a framework for self-supervised learning of speech representa- tions,” in Proceedings of the 34th International Conference on Neural Information Processing Systems , ser. NIPS ’20. Red Hook, NY , USA: Curran Associate...

    2020

  18. [26]

    KenLM: Faster and smaller language model queries,

    K. Heafield, “KenLM: Faster and smaller language model queries,” in Proceedings of the Sixth Workshop on Statistical Ma- chine Translation, C. Callison-Burch, P. Koehn, C. Monz, and O. F. Zaidan, Eds. Edinburgh, Scotland: Association for Com- putational Linguistics, Jul. 2011,...

    2011

  19. [27]

    Neighbourhoods of words in the mental lexicon,

    P. A. Luce, “Neighbourhoods of words in the mental lexicon,” In- diana University, Bloomington, IN, Tech. Rep. Technical Report No. 6, 1986

    1986

  20. [28]

    Binary codes capable of correcting deletions, insertions, and reversals,

    V . I. Levenshtein, “Binary codes capable of correcting deletions, insertions, and reversals,” in Soviet Physics Doklady , vol. 10, no. 8, 1966, pp. 707–710

    1966

  21. [29]

    [Online]

    R Core Team, R: A Language and Environment for Statistical Computing, R Foundation for Statistical Computing, Vienna, Austria, 2023. [Online]. Available: https://www.R-project.org/

    2023

  22. [30]

    An analysis of large language models for African American English speaking children’s oral language assessment,

    A. Johnson, C. Chance, K. Stiemke, H. Veeramani, N. B. Shankar, and A. Alwan, “An analysis of large language models for African American English speaking children’s oral language assessment,” Journal of Black Excellence in Engineering, Science, & Technol- ogy, vol. 1, dec 4 2023

    2023

  23. [31]

    Self-supervised speech representations still struggle with African American Vernacular English,

    K. Chang, Y .-H. Chou, J. Shi, H.-M. Chen, N. Holliday, O. Scharenborg, and D. R. Mortensen, “Self-supervised speech representations still struggle with African American Vernacular English,” in Interspeech 2024, 2024, pp. 4643–4647

    2024

  24. [32]

    Transfer learning from adult to children for speech recognition: Evaluation, analysis and recommendations,

    P. Gurunath Shivakumar and P. Georgiou, “Transfer learning from adult to children for speech recognition: Evaluation, analysis and recommendations,” Computer Speech & Language , vol. 63, p. 101077, 2020

    2020

  25. [33]

    Preprocessing for elderly speech recognition of smart devices,

    S. Kwon, S.-J. Kim, and J. Y . Choeh, “Preprocessing for elderly speech recognition of smart devices,” Computer Speech & Lan- guage, vol. 36, pp. 110–121, 2016

    2016

  26. [34]

    Effects of talker dialect, gender and race on accuracy of Bing Speech and YouTube automatic cap- tions,

    R. Tatman and C. Kasten, “Effects of talker dialect, gender and race on accuracy of Bing Speech and YouTube automatic cap- tions,” in Interspeech 2017, 2017, pp. 934–938

    2017

  27. [35]

    Modeling gender and dialect bias in automatic speech recognition,

    C. Harris, C. Mgbahurike, N. Kumar, and D. Yang, “Modeling gender and dialect bias in automatic speech recognition,” inFind- ings of the Association for Computational Linguistics: EMNLP 2024, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: Association for...

    2024

  28. [36]

    Towards end-to-end speech recognition with recurrent neural networks,

    A. Graves and N. Jaitly, “Towards end-to-end speech recognition with recurrent neural networks,” in Proceedings of the 31st In- ternational Conference on Machine Learning , ser. Proceedings of Machine Learning Research, E. P. Xing and T. Jebara, Eds., vol. 32, no. 2. Bejing, C...

    2014

  29. [37]

    Analysis of LLM as a grammatical feature tagger for African American English,

    R. Porwal, A. Rozet, J. Gowda, P. Houck, K. Tang, and S. Moeller, “Analysis of LLM as a grammatical feature tagger for African American English,” in Findings of the Association for Compu- tational Linguistics: NAACL 2025 , L. Chiruzzo, A. Ritter, and L. Wang, Eds. Albuquerque,...

    2025

Pith tools

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