Pith. sign in

REVIEW 4 major objections 5 minor 83 references

This paper introduces CAMÕES, the first open ASR benchmark for European Portuguese—a 46-hour, five-domain test set—and claims state-of-the-art models with over 35% relative WER improvement over the strongest zero-shot foundation model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

CAMOES is a new open benchmark and model collection for European Portuguese ASR, cutting word error rate by about 35% over the best zero-shot model.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A genuinely useful EP ASR benchmark with credible baselines, but the headline WER numbers need a contamination check before they can be taken at face value. the 4 major comments →

arxiv 2508.19721 v1 pith:Y6HVGOHR submitted 2025-08-27 cs.CL eess.AS

CAM\~OES: A Comprehensive Automatic Speech Recognition Benchmark for European Portuguese

Carlos Carvalho , Francisco Teixeira , Catarina Botelho , Anna Pompili , Rub\'en Solera-Ure\~na , S\'ergio Paulo , Mariana Juli\~ao , Thomas Rolland
show 4 more authors
John Mendon\c{c}a Diogo Pereira Isabel Trancoso Alberto Abad
This is my paper
classification cs.CL eess.AS
keywords automatic speech recognitionEuropean Portuguesebenchmarkfoundation modelsE-BranchformerWhisperlow-resource languagesPortuguese varieties
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

European Portuguese has been largely left out of modern end-to-end speech recognition, which concentrates on Brazilian Portuguese. This paper introduces CAMÕES, a publicly released benchmark made of 46 hours of European Portuguese test speech organized into five domains by spontaneity, plus smaller evaluation sets for African/Asian and Brazilian Portuguese, and a set of open models trained on 425 hours of curated European Portuguese. The paper claims that fine-tuned foundation models, particularly WhisperLv3 with voice-activity-detection preprocessing, beat the strongest zero-shot model by more than 35% relative word error rate, and that a from-scratch 114M-parameter E-Branchformer with a w2v-BERT2 encoder comes within about two WER points of that fine-tuned model. If these claims hold, researchers get the first common yardstick and strong baselines for European Portuguese ASR, and a path to better coverage for other under-resourced Portuguese varieties.

Core claim

The central discovery is that a multi-domain benchmark, assembled from existing corpora plus in-house and crawled data, can support state-of-the-art ASR for under-resourced Portuguese varieties when combined with large multilingual foundation models. The best model—WhisperLv3-X fine-tuned on EP-425—achieves a 12.5% average WER across the five EP domains, a relative improvement of 35% or more over its own zero-shot performance, and outperforms all prior EP systems. Training an E-Branchformer from scratch with frozen w2v-BERT2 features nearly closes the gap to the fine-tuned Whisper, and a single model trained jointly on European, Brazilian, and African/Asian Portuguese matches the variety-spe

What carries the argument

The load-bearing object is the CAMÕES benchmark itself: 46 hours of European Portuguese test speech organized into five domains graded by spontaneity (read speech, broadcast news, talks/lectures, conversational speech, sociolinguistic interviews), plus 3.4 hours of African/Asian Portuguese and 13.2 hours of Brazilian Portuguese, with per-domain averages reported so each domain weighs equally. The models are built on two mechanisms: (1) fine-tuning large supervised foundation models, notably WhisperLv3 with WhisperX voice-activity detection that suppresses hallucination in noisy non-speech segments; and (2) training an E-Branchformer from scratch on FBank features or with SSL encoders (XLSR o

Load-bearing premise

The headline WER numbers depend on the test partitions being clean, representative, and not contaminated by training data—the paper itself notes that in one corpus (Lectra lectures) speakers are shared between train and test, and for automatically aligned audiobooks the reference text may contain alignment errors.

What would settle it

Run the released best models on a new, carefully speaker-disjoint European Portuguese test set (e.g., a held-out corpus not used in training) and compare the WER to the reported 12.5% EP average; a substantial jump would indicate the benchmark numbers are inflated by speaker overlap or alignment leakage. A simpler audit: check all 14 test corpora for speaker-ID overlap with the 425-hour training set beyond the acknowledged Lectra case.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The 46-hour, five-domain test set gives European Portuguese a reproducible yardstick; any future ASR system can be scored against the released models and the leaderboard.
  • Fine-tuned WhisperLv3-X, with VAD preprocessing, is the strongest EP model overall, and the from-scratch E-Branchformer + w2v-BERT2 at 114M trainable parameters reaches within ~2% WER of it, suggesting SSL features can substitute for large supervised pretraining.
  • Joint training on EP+BP+AAP yields a single model that matches variety-specific models on all varieties, pointing toward one multi-variety Portuguese system rather than separate ones.
  • The benchmark's per-domain breakdown shows sociolinguistic interviews remain the hardest domain, with WERs around 27% even after fine-tuning.
  • For Brazilian Portuguese, the best models (WhisperLv3-X and E-Branchformer-w2vBERT2, whether BP-only or PT-All) outperform prior SOTA on CORAA and MuPe test sets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the paper reports per-domain averages, the benchmark is immediately usable to diagnose where foundation models fail: the SI domain's high WERs suggest that spontaneous, accented, low-quality conversational speech is the frontier for low-resource ASR.
  • The observation that fine-tuning on EP degrades BP performance and vice versa, while joint training preserves both, implies that a single Portuguese model trained on all varieties may generalize to other Lusophone accents (Angola, Mozambique, East Timor, Goa, Macao) better than a European-only model—a testable extension given the benchmark's AAP set.
  • The prompt-tuning result for Phi-4-MI—where specifying 'European Portuguese' in the prompt lowers WER zero-shot—suggests that dialect-aware prompting is a cheap, zero-parameter way to improve multimodal LLM ASR, and it may carry over to other dialect/language pairs.
  • If the benchmark is adopted, its current contamination caveat (shared speakers in the Lectra talks/lectures partition) will need to be addressed or annotated, since inflated lecture-domain numbers could mislead downstream comparisons.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces CAMÕES, an open ASR benchmark and model collection for European Portuguese (EP) and other Portuguese varieties. The benchmark comprises 46.2 h of EP test audio across five domains (read speech, broadcast news, talks/lectures, conversational speech, sociolinguistic interviews), plus smaller evaluation sets for African/Asian Portuguese (AAP) and Brazilian Portuguese (BP). The authors curate 425 h of EP training data (EP-425), evaluate several zero-shot foundation models (Whisper Large v3, OWSM-CTC, MMS, SeamlessM4T, Phi-4-MI), fine-tune the two strongest (WhisperLv3 and Phi-4-MI) on EP-425, and train E-Branchformer models with FBank or SSL features. They report large WER reductions from fine-tuning—e.g., WhisperLv3-X goes from 16.4% to 7.4% average EP WER—and claim relative improvements above 35% over the strongest zero-shot foundation model, establishing a new SOTA for EP and other Portuguese varieties. They also evaluate variety-specific and multi-variety models on BP and AAP, including a comparison with prior BP SOTA on CORAA and MuPe.

Significance. If the benchmark is clean and the results are reproducible, this is a valuable contribution: it is the first open, comprehensive EP ASR benchmark, it releases trained models and a leaderboard, and it provides a realistic picture of zero-shot and fine-tuned performance for modern foundation models on a low-resource variety. The systematic evaluation across five domains and the inclusion of demographic analyses are also useful. The paper ships concrete artifacts (Hugging Face datasets/models), and the central WER numbers are internally consistent. The main value lies in becoming a reference evaluation suite for EP ASR, provided the test partitions are shown to be uncontaminated and speaker-disjoint.

major comments (4)
  1. [Table I, Section III-A] The test-set integrity is load-bearing for the headline SOTA claims, and Table I admits that the Lectra corpus shares all 7 speakers between the 22.0 h training partition and the 2.6 h test partition. Since Lectra accounts for 2.6 of the 3.0 h in the T/L domain, the fine-tuned T/L WER of 12.5% (vs. 19.2% zero-shot) may partly reflect speaker memorization rather than generalization. The paper does not quantify the impact, nor does it state speaker disjointness for the other corpora with both train and test partitions (BD-Publico, PT Adults, PT Elderly, SpeechDat, Alert, MuA ViC, Postport, Fala Bracarense, MLS extended). I request a speaker-overlap analysis for every such corpus and, where overlap exists, either a re-split into disjoint speaker sets or a per-speaker held-out evaluation. This is necessary before the reported improvements and SOTA claims can be accepted.
  2. [Section IV-A, Table I] Several test components are derived from public datasets (CommonVoice, VoxCelebPT, MLS extended) that are plausible constituents of the pretraining corpora of Whisper, w2v-BERT2, XLSR, and other foundation models. No contamination analysis is provided. If these test utterances (or near-duplicates) appeared in pretraining, the zero-shot baselines and the fine-tuned gains could both be inflated, undermining the comparison with the strongest zero-shot model and the claimed state of the art. The authors should either check and report overlap (e.g., via audio fingerprinting or utterance-ID matching) or clearly restrict the benchmark to portions verified to be unseen by the evaluated models.
  3. [Section VI-A, prompt selection] The Phi-4-MI zero-shot results are obtained after selecting among three prompts based on their performance on the CAMÕES evaluation set. This is a form of test-set tuning: the reported 21.7% average EP WER is the best of three prompt choices, not a genuine zero-shot measurement. It also makes the Phi-4-MI comparison with other models unfair, since the other models did not receive a comparable prompt-selection procedure. I recommend selecting prompts on a separate held-out validation set (or reporting all prompt variants) and clearly marking the results as prompt-tuned rather than zero-shot.
  4. [Section III-A, Table I (MLS extended)] The MLS extended test set is transcribed by automatic alignment at book or chapter level. The paper acknowledges this, but the effect of reference-label noise on the reported WER is not analyzed. With only 1.0 h of test audio the impact on the overall benchmark is limited, but for a benchmark whose purpose is precise model comparison, it would be prudent to either manually verify a sample of those alignments or remove this subset from the headline numbers and report it separately.
minor comments (5)
  1. [Table III] The column headers are not self-explanatory in the text; the reader must infer that the first three numeric columns are EP, AAP, BP and the next five are RS, BN, T/L, CS, SI, Avg. Please add explicit column labels in the table caption or a separate header row.
  2. [Section IV-A] The benchmark average is computed with equal domain weights, but the domains have very different test durations (from 0.4 h for MuA ViC to 9.7 h for SpeechDat). This weighting choice should be stated explicitly in the benchmark definition, and a duration-weighted average should be reported alongside.
  3. [Section V] The text says E-Branchformer models are 'trained from scratch' but then uses frozen SSL encoders (XLSR or w2v-BERT2) as feature extractors. This is a meaningful distinction; consider using 'trained from scratch on the ASR head' or similar.
  4. [Section VI-A] The WhisperX VAD parameters are mentioned as a factor in the WhisperLv3-X results, but no details or sensitivity analysis are provided. Since the paper repeatedly contrasts WhisperLv3 with WhisperLv3-X, the VAD configuration should be described precisely.
  5. [References] Reference [69] contains a typo ('olthers' instead of 'others'); please correct.

Circularity Check

0 steps flagged

No circularity: the paper's SOTA claims are empirical measurements on held-out test partitions, not derivations that reduce to their own inputs.

full rationale

This paper contains no derivation chain that could be circular: CAMÕES is an evaluation benchmark plus trained models, and the headline numbers (e.g., 19.2% zero-shot vs 12.5% fine-tuned WhisperLv3-X WER, 'relative improvements above 35%') are direct empirical measurements on held-out test partitions, not quantities derived by construction from the training data or from an assumed model. The test set is not defined in terms of the models' outputs, and no fitted parameter is renamed as a prediction. The self-citations (e.g., [19], [41]-[43], [52], [58]-[63], [76]) are used for corpus provenance, prior context, or training-schedule details; none is invoked as a uniqueness theorem or as the justification for the central benchmark/SOTA claim. The only validity concerns are benchmark-quality issues, not circularity: Table I discloses that Lectra speakers are shared between train and test, MLS extended uses book/chapter-level automatic alignment, and prompt/model selection used the evaluation set; these can inflate or tune results but do not make the reported WERs true by construction. Against external benchmarks (CORAA, MuPe), the comparison is independent of the paper's own fits. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The paper introduces no new theoretical entities. It relies on standard ASR methodology and existing corpora. The main unstated supports are data-quality assumptions about the curated test set and the fairness of the WER metric, both of which are partially documented.

free parameters (3)
  • Phi-4-MI prompt selection = "Transcribe the European Portuguese audio clip into text"
    The best prompt was chosen based on zero-shot performance on the evaluation set (Section VI.A), which is a tuning parameter fitted to the test data.
  • Equal domain weighting in benchmark average = 1/5 per domain
    The aggregate Avg. column weights each of the five domains equally even though domain hours differ (Section IV.A). This choice affects the headline numbers.
  • WhisperX VAD parameters = Not specified
    WhisperLv3-X relies on WhisperX voice activity detection; the specific VAD configuration is not reported, so its contribution to the WER improvement is not independently tunable.
axioms (3)
  • domain assumption The curated train and test partitions are speaker-disjoint and the transcriptions are accurate.
    Section III.A describes corpus curation but does not verify speaker disjointness for all corpora; Table I itself flags Lectra as having shared speakers between partitions, and spoken books use automatic alignment at book/chapter level, so this assumption is partially violated.
  • domain assumption WER computed with the Whisper-based text normalizer is a fair and consistent metric across all models.
    Section V states that all models use the same text normalizer based on Whisper, which standardizes comparison but may favor models whose outputs resemble Whisper's normalized style.
  • domain assumption The five CAMOES domains are representative of European Portuguese speech across demographics and recording conditions.
    Section IV.A asserts the test set spans age groups, regions, and spontaneity levels, but the actual distribution is uneven (e.g., only 0.4h of TEDx talks in MuA ViC), so representativeness is an assertion rather than a proven property.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CAM\~OES: A Comprehensive Automatic Speech Recognition Benchmark for European Portuguese." pith.science (2026). https://pith.science/paper/Y6HVGOHR

@misc{pith2026250819721,
  author       = {Pith},
  title        = {Pith review of: CAM\~OES: A Comprehensive Automatic Speech Recognition Benchmark for European Portuguese},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y6HVGOHR}},
  note         = {Machine review of arXiv:2508.19721}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing resources for Automatic Speech Recognition in Portuguese are mostly focused on Brazilian Portuguese, leaving European Portuguese (EP) and other varieties under-explored. To bridge this gap, we introduce CAM\~OES, the first open framework for EP and other Portuguese varieties. It consists of (1) a comprehensive evaluation benchmark, including 46h of EP test data spanning multiple domains; and (2) a collection of state-of-the-art models. For the latter, we consider multiple foundation models, evaluating their zero-shot and fine-tuned performances, as well as E-Branchformer models trained from scratch. A curated set of 425h of EP was used for both fine-tuning and training. Our results show comparable performance for EP between fine-tuned foundation models and the E-Branchformer. Furthermore, the best-performing models achieve relative improvements above 35% WER, compared to the strongest zero-shot foundation model, establishing a new state-of-the-art for EP and other varieties.

Figures

Figures reproduced from arXiv: 2508.19721 by Alberto Abad, Anna Pompili, Carlos Carvalho, Catarina Botelho, Diogo Pereira, Francisco Teixeira, Isabel Trancoso, John Mendon\c{c}a, Mariana Juli\~ao, Rub\'en Solera-Ure\~na, S\'ergio Paulo, Thomas Rolland.

Figure 1
Figure 1. Figure 1: WER [%] performance for WhisperLv3-X zero-shot and fine-tuned, per age range (a) and per gender (b), on the RS domain (BD-Publico, PT Adults, PT Children, PT Elderly, and SpeechDat). The figure also shows the number of hours per age range and gender considered for this analysis. primarily trained on BP. Similarly, EBranch models trained from scratch with EP speech also show improvements for AAP, with resul… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

83 extracted references · 75 canonical work pages · 1 internal anchor

  1. [1]

    The most spoken languages worldwide in 2025

    Statista, “The most spoken languages worldwide in 2025.” https://www. statista.com/statistics/266808/the-most-spoken-languages-worldwide/,

    2025

  2. [2]

    Portuguese language

    Ethnologue, SIL International, “Portuguese language.” https://www. ethnologue.com/language/por/, n.d. Accessed: May 29, 2025

    2025

  3. [3]

    M. H. Mateus and E. d’Andrade, The Phonology Of Portuguese . Oxford University Press, 2000

    work page 2000

  4. [4]

    Common V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila et al. , “Common V oice: A Massively-Multilingual Speech Corpus,” in Proc. LREC, pp. 4218–4222, 2020

    work page 2020

  5. [5]

    CORAA ASR: a large corpus of spontaneous and prepared speech manually validated for speech recognition in Brazilian Portuguese,

    A. Candido Junior et al., “CORAA ASR: a large corpus of spontaneous and prepared speech manually validated for speech recognition in Brazilian Portuguese,” Language Resources and Evaluation , vol. 57, pp. 1139–1171, 2023

    work page 2023

  6. [6]

    FLEURS: Few-shot Learning Evaluation of Univer- sal Representations of Speech,

    A. Conneau et al., “FLEURS: Few-shot Learning Evaluation of Univer- sal Representations of Speech,” in Proc. SLT, pp. 798–805, 2023

    work page 2023

  7. [7]

    Yodas: Youtube-Oriented Dataset for Audio and Speech,

    X. Li et al. , “Yodas: Youtube-Oriented Dataset for Audio and Speech,” in Proc. ASRU, pp. 1–8, 2023

    work page 2023

  8. [8]

    Cem Mil Podcasts: A Spoken Portuguese Document Corpus for Multi-modal, Multi-lingual and Multi-dialect Information Access Research,

    E. Garmash et al., “Cem Mil Podcasts: A Spoken Portuguese Document Corpus for Multi-modal, Multi-lingual and Multi-dialect Information Access Research,” in Proc. Experimental IR Meets Multilinguality, Multimodality, and Interaction , pp. 48–59, 2023

    work page 2023

  9. [9]

    Speech-Transformer: A No-Recurrence Sequence-to-Sequence Model for Speech Recognition,

    L. Dong, S. Xu, and B. Xu, “Speech-Transformer: A No-Recurrence Sequence-to-Sequence Model for Speech Recognition,” in Proc. ICASSP, pp. 5884–5888, 2018

    work page 2018

  10. [10]

    A Comparative Study on Transformer vs RNN in Speech Applications,

    S. Karita et al. , “A Comparative Study on Transformer vs RNN in Speech Applications,” in Proc. ASRU, pp. 449–456, 2019

    work page 2019

  11. [11]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati et al. , “Conformer: Convolution-augmented transformer for speech recognition,” in Proc. Interspeech, pp. 5036–5040, 2020

    work page 2020

  12. [12]

    E-Branchformer: Branchformer with Enhanced Merging for Speech Recognition,

    K. Kim et al., “E-Branchformer: Branchformer with Enhanced Merging for Speech Recognition,” in Proc. SLT, pp. 84–91, 2023

    work page 2023

  13. [13]

    Fast Conformer With Linearly Scalable Attention For Efficient Speech Recognition,

    D. Rekesh et al., “Fast Conformer With Linearly Scalable Attention For Efficient Speech Recognition,” in Proc. ASRU, pp. 1–8, 2023

    work page 2023

  14. [14]

    Speechstew: Simply mix all available speech recognition data to train one large neural network,

    W. Chan et al. , “Speechstew: Simply mix all available speech recognition data to train one large neural network,” arXiv preprint arXiv:2104.02133, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    Robust speech recognition via large-scale weak supervision,

    A. Radford et al. , “Robust speech recognition via large-scale weak supervision,” in Proc. ICML, 2023

    work page 2023

  16. [16]

    Reproducing Whisper-Style Training Using An Open- Source Toolkit And Publicly Available Data,

    Y . Peng et al. , “Reproducing Whisper-Style Training Using An Open- Source Toolkit And Publicly Available Data,” in Proc. ASRU, pp. 1–8, 2023

    work page 2023

  17. [17]

    OWSM-CTC: An Open Encoder-Only Speech Foundation Model for Speech Recognition, Translation, and Language Identification,

    Y . Peng, Y . Sudo, M. Shakeel, and S. Watanabe, “OWSM-CTC: An Open Encoder-Only Speech Foundation Model for Speech Recognition, Translation, and Language Identification,” in Proc. ACL (V olume 1: Long Papers), pp. 10192–10209, 2024

    work page 2024

  18. [18]

    A corpus-based study of elderly and young speakers of European Portuguese: acoustic correlates and their impact on speech recognition performance,

    T. Pellegrini et al., “A corpus-based study of elderly and young speakers of European Portuguese: acoustic correlates and their impact on speech recognition performance,” in Proc. Interspeech, pp. 852–856, 2013

    work page 2013

  19. [19]

    TRIBUS: An end-to-end automatic speech recognition system for European Portuguese,

    C. Carvalho and A. Abad, “TRIBUS: An end-to-end automatic speech recognition system for European Portuguese,” in Proc. IberSPEECH , pp. 185–189, 2021

    work page 2021

  20. [20]

    Scaling Speech Technology to 1,000+ Languages,

    V . Pratap et al., “Scaling Speech Technology to 1,000+ Languages,” The Journal of Machine Learning Research , vol. 25, no. 1, pp. 4798–4849, 2024

    work page 2024

  21. [21]

    Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages,

    Y . Zhang et al. , “Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages,” arXiv preprint arXiv:2303.01037 , 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Seamless: Multilingual Expressive and Streaming Speech Translation,

    L. Barrault et al. , “Seamless: Multilingual Expressive and Streaming Speech Translation,” arXiv preprint arXiv:2312.05187 , 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Towards Robust Speech Representation Learning for Thousands of Languages,

    W. Chen et al. , “Towards Robust Speech Representation Learning for Thousands of Languages,” in Proc. EMNLP , pp. 10205–10224, Nov. 2024

    work page 2024

  24. [24]

    Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs,

    A. Abouelenin et al. , “Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs,” arXiv preprint arXiv:2503.01743, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    OWSM v4: Improving Open Whisper-Style Speech Models via Data Scaling and Cleaning,

    Y . Peng et al. , “OWSM v4: Improving Open Whisper-Style Speech Models via Data Scaling and Cleaning,” in Proc. Interspeech (accepted), 2025

    work page 2025

  26. [26]

    ML-SUPERB: Multilingual Speech Universal PERfor- mance Benchmark,

    J. Shi et al. , “ML-SUPERB: Multilingual Speech Universal PERfor- mance Benchmark,” in Proc. Interspeech, pp. 884–888, 2023

    work page 2023

  27. [27]

    SUPERB: Speech Processing Universal PERfor- mance Benchmark,

    S.-W. Yang et al. , “SUPERB: Speech Processing Universal PERfor- mance Benchmark,” in Proc. Interspeech 2021 , pp. 1194–1198, 2021

    work page 2021

  28. [28]

    WER We Stand: Benchmarking Urdu ASR Models,

    S. Arif et al. , “WER We Stand: Benchmarking Urdu ASR Models,” in Proc. International Conference on Computational Linguistics , pp. 5952– 5961, 2025

    work page 2025

  29. [29]

    Comprehensive Benchmarking and Analysis of Open Pretrained Thai Speech Recognition Models,

    P. Tipakasorn et al. , “Comprehensive Benchmarking and Analysis of Open Pretrained Thai Speech Recognition Models,” in Proc. Confer- ence of the Oriental COCOSDA International Committee for the Co- ordination and Standardisation of Speech Databases and Assessment Techniques (O-COCOSDA), pp. 1–7, 2024

    work page 2024

  30. [30]

    Speech Recognition for Greek Dialects: A Chal- lenging Benchmark,

    S. Vakirtzian et al. , “Speech Recognition for Greek Dialects: A Chal- lenging Benchmark,” in Proc. Interspeech, pp. 3974–3978, 2024

    work page 2024

  31. [31]

    Benchmarking Whisper for Low-Resource Speech Recognition: An N-Shot Evaluation on Pashto, Punjabi, and Urdu,

    N. U. Sehar et al. , “Benchmarking Whisper for Low-Resource Speech Recognition: An N-Shot Evaluation on Pashto, Punjabi, and Urdu,” in Proc. First Workshop on Challenges in Processing South Asian Languages (CHiPSAL) , pp. 202–207, 2025

    work page 2025

  32. [32]

    The Development of a Speaker Independent Continuous Speech Recognizer for Portuguese,

    J. P. Neto, C. A. Martins, and L. B. Almeida, “The Development of a Speaker Independent Continuous Speech Recognizer for Portuguese,” in Proc. Eurospeech, pp. 1703–1706, 1997

    work page 1997

  33. [33]

    Feature Sets in Continuous Speech Recognition for the Portuguese Language,

    S. Dos Santos and A. Alcaim, “Feature Sets in Continuous Speech Recognition for the Portuguese Language,” in Proc. IEEE International Telecommunications Symposium, pp. 126–129, 1998. Aug 9–13, 1998

    work page 1998

  34. [34]

    Speech recognition of broadcast news for the European Portuguese language,

    H. Meinedo, N. Souto, and J. Neto, “Speech recognition of broadcast news for the European Portuguese language,” in Proc. ASRU, pp. 319– 322, 2001

    work page 2001

  35. [35]

    Broadcast news subtitling system in Portuguese,

    J. Neto et al., “Broadcast news subtitling system in Portuguese,” in Proc. ICASSP, pp. 1561–1564, 2008

    work page 2008

  36. [36]

    Reconhecimento de fala em portugu ˆes de Portugal num contexto com poucos recursos,

    J. M. A. M. de S ´a, “Reconhecimento de fala em portugu ˆes de Portugal num contexto com poucos recursos,” master’s thesis, Faculdade de Ciˆencias, Universidade do Porto, Porto, Portugal, November 2021. Available at https://hdl.handle.net/10216/139258

    work page 2021

  37. [37]

    Deep learning for speech to text transcription for the Portuguese language,

    E. F. Medeiros, “Deep learning for speech to text transcription for the Portuguese language,” master’s thesis, Escola de Ci ˆencias e Tecnologia, Universidade de ´Evora, ´Evora, Portugal, February 2023. Available at http://hdl.handle.net/10174/34859

    work page 2023

  38. [38]

    Domain Adaptation Speech-to-Text for Low- Resource European Portuguese Using Deep Learning,

    E. Medeiros et al. , “Domain Adaptation Speech-to-Text for Low- Resource European Portuguese Using Deep Learning,” Future Internet, vol. 15, no. 5, 2023

    work page 2023

  39. [39]

    Desenvolvimento de um Sistema de Reconhecimento Autom´atico de V oz Cont ´ınua com Grande V ocabul ´ario para o Por- tuguˆes Brasileiro,

    E. Silva et al. , “Desenvolvimento de um Sistema de Reconhecimento Autom´atico de V oz Cont ´ınua com Grande V ocabul ´ario para o Por- tuguˆes Brasileiro,” in Proc. XXV Congresso da Sociedade Brasileira de Computac ¸˜ao, pp. 2258–2267, 2005

    work page 2005

  40. [40]

    Free tools and resources for Brazilian Portuguese speech recognition,

    N. Neto et al., “Free tools and resources for Brazilian Portuguese speech recognition,” Journal of the Brazilian Computer Society , vol. 17, pp. 53– 68, 11 2010

    work page 2010

  41. [41]

    Porting an European Portuguese broadcast news recog- nition system to Brazilian Portuguese,

    A. Abad et al., “Porting an European Portuguese broadcast news recog- nition system to Brazilian Portuguese,” in Proc. Interspeech, pp. 92–95, 2009

    work page 2009

  42. [42]

    Exploiting variety-dependent phones in portuguese variety identification applied to broadcast news transcription,

    O. Koller et al. , “Exploiting variety-dependent phones in portuguese variety identification applied to broadcast news transcription,” in Proc. Interspeech, pp. 749–752, 2010

    work page 2010

  43. [43]

    Transcription of Multi-variety Portuguese Media Con- tents,

    A. Abad et al. , “Transcription of Multi-variety Portuguese Media Con- tents,” in Proc. PROPOR, pp. 409–420, 2012

    work page 2012

  44. [44]

    LSF and LPC-derived features for large vocabulary distributed continuous speech recognition in Brazilian Por- tuguese,

    V . Alencar and A. Alcaim, “LSF and LPC-derived features for large vocabulary distributed continuous speech recognition in Brazilian Por- tuguese,” in Proc. 42nd Asilomar Conference on Signals, Systems and Computers, pp. 1237–1241, 2008

    work page 2008

  45. [45]

    A Large Dataset of Spontaneous Speech with the Accent Spoken in S ˜ao Paulo for Automatic Speech Recognition Evaluation,

    R. Lima et al., “A Large Dataset of Spontaneous Speech with the Accent Spoken in S ˜ao Paulo for Automatic Speech Recognition Evaluation,” in Proc. Intelligent Systems: 34th Brazilian Conference (BRACIS) , pp. 33– 47, 2025

    work page 2025

  46. [46]

    MuPe Life Stories Dataset: Spontaneous Speech in Brazilian Portuguese with a Case Study Evaluation on ASR Bias against Speakers Groups and Topic Modeling,

    S. Evaldo Leal et al. , “MuPe Life Stories Dataset: Spontaneous Speech in Brazilian Portuguese with a Case Study Evaluation on ASR Bias against Speakers Groups and Topic Modeling,” in Proc. 31st Interna- tional Conference on Computational Linguistics , pp. 6076–6087, 2025

    work page 2025

  47. [47]

    Baseline acoustic models for Brazilian Portuguese using Kaldi tools,

    C. Batista, A. L. Dias, and N. Sampaio Neto, “Baseline acoustic models for Brazilian Portuguese using Kaldi tools,” in IberSPEECH 2018 , pp. 77–81, 2018

    work page 2018

  48. [48]

    Brazilian Portuguese Speech Recognition Using Wav2vec 2.0,

    L. R. Stefanel Gris et al. , “Brazilian Portuguese Speech Recognition Using Wav2vec 2.0,” in Proc. PROPOR, pp. 333–343, 2022

    work page 2022

  49. [49]

    LLM based Text Generation for Improved Low- resource Speech Recognition Models,

    T. Nagano et al. , “LLM based Text Generation for Improved Low- resource Speech Recognition Models,” in Proc. ICASSP, pp. 1–5, 2025

    work page 2025

  50. [50]

    The design of a large vocabulary speech corpus for portuguese,

    J. P. Neto et al. , “The design of a large vocabulary speech corpus for portuguese,” in Proc. Eurospeech, pp. 1707–1710, 1997

    work page 1997

  51. [51]

    The DIRHA Portuguese Corpus: A Comparison of Home Automation Command Detection and Recognition in Simulated and Real Data,

    M. Matos, A. Abad, and A. Serralheiro, “The DIRHA Portuguese Corpus: A Comparison of Home Automation Command Detection and Recognition in Simulated and Real Data,” in Proc. LREC , pp. 4012– 4018, 2016

    work page 2016

  52. [52]

    Paulo, Automatic Methods for Building Speech Synthesis Corpora

    S. Paulo, Automatic Methods for Building Speech Synthesis Corpora . PhD thesis, Instituto Superior T ´ecnico, Universidade de Lisboa, Lisboa, Portugal, June 2009. Available at http://

    work page 2009

  53. [53]

    MLS: A Large-Scale Multilingual Dataset for Speech Research,

    V . Pratap et al. , “MLS: A Large-Scale Multilingual Dataset for Speech Research,” in Proc. Interspeech, pp. 2757–2761, 2020

    work page 2020

  54. [54]

    Improving speech recognition through automatic selection of age group–specific acoustic models,

    A. H ¨am¨al¨ainen et al., “Improving speech recognition through automatic selection of age group–specific acoustic models,” in Proc. PROPOR , pp. 12–23, Springer, 2014

    work page 2014

  55. [55]

    The CNG corpus of European Portuguese children’s speech,

    A. H ¨am¨al¨ainen et al. , “The CNG corpus of European Portuguese children’s speech,” in Proc. International Conference on Text, Speech and Dialogue , pp. 544–551, 2013

    work page 2013

  56. [56]

    The first European Portuguese elderly speech corpus,

    A. H ¨am¨al¨ainen et al. , “The first European Portuguese elderly speech corpus,” Proc. IberSPEECH, vol. 10, 2012

    work page 2012

  57. [57]

    HMM/MLP hybrid speech recognizer for the Portuguese telephone SpeechDat corpus,

    A. Hagen and J. P. Neto, “HMM/MLP hybrid speech recognizer for the Portuguese telephone SpeechDat corpus,” in Proc. PROPOR, pp. 126– 134, 2003

    work page 2003

  58. [58]

    Evaluation of an alert system for selective dissemi- nation of broadcast news,

    I. Trancoso et al. , “Evaluation of an alert system for selective dissemi- nation of broadcast news,” in Proc. Interspeech, pp. 1257–1260, 2003

    work page 2003

  59. [59]

    The LECTRA corpus – classroom lecture transcrip- tions in European Portuguese,

    I. Trancoso et al. , “The LECTRA corpus – classroom lecture transcrip- tions in European Portuguese,” in Proc. LREC, pp. 1416–1420, 2008

    work page 2008

  60. [60]

    MuA ViC: A Multilingual Audio-Visual Corpus for Robust Speech Recognition and Robust Speech-to-Text Translation,

    M. Anwar et al. , “MuA ViC: A Multilingual Audio-Visual Corpus for Robust Speech Recognition and Robust Speech-to-Text Translation,” in Proc. Interspeech, pp. 4064–4068, 2023

    work page 2023

  61. [61]

    Corpus de di ´alogo CORAL,

    I. Trancoso et al., “Corpus de di ´alogo CORAL,” Proc. PROPOR, 1998

    work page 1998

  62. [62]

    The L2F broadcast news speech recognition system,

    H. Meinedo et al., “The L2F broadcast news speech recognition system,” in Proc. F ALA 2010, pp. 93–96, 2010

    work page 2010

  63. [63]

    V oxCeleb-PT - a dataset for a speech processing course,

    J. Mendonc ¸a and I. Trancoso, “V oxCeleb-PT - a dataset for a speech processing course,” in Proc. IberSPEECH 2022 , pp. 71–75, 2022

    work page 2022

  64. [64]

    Per- fil Sociolingu ´ıstico da Fala Bracarense

    Centro de Estudos Human ´ısticos, Universidade do Minho, “Per- fil Sociolingu ´ıstico da Fala Bracarense.” https://sites.google.com/site/ projectofalabracarense/, 2009. Accessed: 2025-05-21

    work page 2009

  65. [65]

    Portugu ˆes Fundamen- tal

    Centro de Lingu ´ıstica, Universidade de Lisboa, “Portugu ˆes Fundamen- tal.” https://www.islrn.org/resources/812-337-422-842-3/, 2014. Ac- cessed: 2025-05-21

    work page 2014

  66. [66]

    Portuguese variety identification on broadcast news,

    J.-L. Rouas et al., “Portuguese variety identification on broadcast news,” in Proc. ICASSP, pp. 4229–4232, 2008

    work page 2008

  67. [67]

    Portugu ˆes Falado: variedades geogr ´aficas e sociais,

    J. Bettencourt Gonc ¸alves, “Portugu ˆes Falado: variedades geogr ´aficas e sociais,” Estudos de gram ´atica portuguesa (1) , pp. 257–266, 2000

    work page 2000

  68. [68]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski et al. , “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Proc. NIPS, vol. 33, pp. 12449– 12460, 2020

    work page 2020

  69. [69]

    WhisperX: Time-Accurate Speech Transcription of Long-Form Audio,

    M. Bain and olthers, “WhisperX: Time-Accurate Speech Transcription of Long-Form Audio,” in Proc. Interspeech, pp. 4489–4493, 2023

    work page 2023

  70. [70]

    Xls-r: Self-supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, et al., “Xls-r: Self-supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021

    Pith/arXiv arXiv 2021

  71. [71]

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

    C. Wang et al., “V oxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpre- tation,” in Proc. of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (V olume 1: Long Papers) , pp. 993– 1003, Aug. 2021

    work page 2021

  72. [72]

    W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,

    Y .-A. Chung et al. , “W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training,” in Proc. ASRU, pp. 244–250, 2021

    work page 2021

  73. [73]

    Espnet: End-to-end speech processing toolkit,

    S. Watanabe et al. , “Espnet: End-to-end speech processing toolkit,” in Proc. Interspeech, pp. 2207–2211, 2018

    work page 2018

  74. [74]

    RoFormer: Enhanced transformer with Rotary Position Embedding,

    J. Su et al. , “RoFormer: Enhanced transformer with Rotary Position Embedding,” Neurocomputing, vol. 568, p. 127063, 2024

    work page 2024

  75. [75]

    Benchmarking Rotary Position Embeddings for Automatic Speech Recognition

    S. Zhang et al. , “Benchmarking Rotary Position Embeddings for Auto- matic Speech Recognition,” arXiv preprint arXiv:2501.06051 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  76. [76]

    Exploring Linear Variant Transformers and k- NN Memory Inference for Long-Form ASR,

    C. Carvalho et al. , “Exploring Linear Variant Transformers and k- NN Memory Inference for Long-Form ASR,” in Proc. Interspeech (accepted), 2025

    work page 2025

  77. [77]

    Europarl: A Parallel Corpus for Statistical Machine Transla- tion,

    P. Koehn, “Europarl: A Parallel Corpus for Statistical Machine Transla- tion,” in Proc. Machine Translation Summit X: Papers , pp. 79–86, 2005

    work page 2005

  78. [78]

    OpenSubtitles2016: Extracting Large Paral- lel Corpora from Movie and TV Subtitles,

    P. Lison and J. Tiedemann, “OpenSubtitles2016: Extracting Large Paral- lel Corpora from Movie and TV Subtitles,” in Proc. LREC, pp. 923–929, May 2016

    work page 2016

  79. [79]

    KenLM: Faster and smaller language model queries,

    K. Heafield, “KenLM: Faster and smaller language model queries,” in Proc. Sixth Workshop on Statistical Machine Translation , pp. 187–197, 2011

    work page 2011

  80. [80]

    Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio,

    M. Bara ´nski et al. , “Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio,” in Proc. ICASSP, pp. 1–5, 2025

    work page 2025

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.