Pith. sign in

REVIEW 4 major objections 5 minor 47 references

A Practitioner's Guide to Building ASR Models for Low-Resource Languages: A Case Study on Scottish Gaelic

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

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

Pith's one-line read A hybrid HMM recognizer beats fine-tuned end-to-end models by 32% on Scottish Gaelic.

desk verdict A useful hybrid-ASR recipe for Gaelic with a clean ablation story, but the headline 32% gain rests on a data-leakage ambiguity the authors need to resolve. read the letter →

arxiv 2506.04915 v1 pith:VR3PRW64 submitted 2025-06-05 cs.CL eess.AS

Ondřej Klejch , William Lamb , Peter Bell This is my paper
classification cs.CLeess.AS
keywords automaticspeechrecognitionlow-resourcelanguagesScottishGaelichybridHMMself-supervisedmodelssemi-supervisedtrainingbyte-pairencodingacousticunitslanguagemodelrescoring
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 overturn a common default in low-resource speech recognition: fine-tuning a large public multilingual end-to-end model is not necessarily the best path. Working on Scottish Gaelic, the authors build a hybrid HMM recognizer whose neural acoustic model runs on features from a self-supervised speech model, uses byte-pair-encoding subword units instead of graphemes, and is rescored by language models trained on all available Gaelic text. This system reaches a 12.8% average word error rate, 32% relative better than their best fine-tuned end-to-end baseline and 54% relative better than the previous Gaelic system. The result matters because many low-resource languages have more untranscribed audio and raw text than transcribed speech, and this recipe is designed to use exactly those resources.

What carries the argument

The load-bearing mechanism is the hybrid HMM pipeline: a neural acoustic model trained with sequence-level maximum mutual information on top of features extracted from a large self-supervised speech model (one trained on unlabeled audio by predicting masked time steps) that was continued-pretrained on target-language audio; byte-pair-encoding subword units as acoustic outputs, which handle Gaelic-English code-switching better than graphemes; subword n-gram and RNN language models trained on all available text and used for lattice rescoring; and two ways of consuming untranscribed audio, continued self-supervised pre-training and pseudo-labeled semi-supervised training, with noise augmentation throughout. The gain is not from any single component but from combining all of them.

What would settle it

Check the broadcast dates in the 184 hours of untranscribed news audio used for semi-supervised training and test whether any of the seven News test episodes (16-22 October 2023) appear; if any do, remove that overlap and recompute the News WER and the overall average.

Watch

Extended reading notes

Core claim

The paper's central claim is that a carefully assembled hybrid model can outperform fine-tuned end-to-end models under limited transcribed data. The authors show that the hybrid's advantage comes from being able to exploit two resources end-to-end fine-tuning does not easily reach: all available text, through subword n-gram and RNN language models, and all available audio, through continued self-supervised pre-training and semi-supervised pseudo-labeling of untranscribed news broadcasts. On Scottish Gaelic the complete recipe gives a 12.8% average WER, a 32% relative improvement over the strongest fine-tuned end-to-end baseline (19.0% WER) and 54% relative over the prior best Gaelic model (28.0% WER achieved by the deployed system). The authors read these numbers as evidence for their broader claim: for low-resource languages with similar amounts of speech and text, the hybrid recipe is a better default than fine-tuning a public multilingual end-to-end model.

Load-bearing premise

The headline comparison is valid only if the 184 hours of untranscribed news audio used for semi-supervised training do not include the seven News test episodes broadcast from 16 to 22 October 2023; the paper explicitly holds out one test set but never says this for News.

Editorial extensions

If this is right

  • For low-resource languages with similar amounts of transcribed speech, untranscribed audio, and raw text, the tuned hybrid recipe is a better starting point than fine-tuning a large public end-to-end model.
  • Continued self-supervised pre-training on target-language audio improves self-supervised features even when the base model saw almost none of the target language.
  • Replacing grapheme acoustic units with BPE subword units reduces word error on code-switched speech.
  • Semi-supervised pseudo-labeling of untranscribed broadcast audio, plus noise augmentation, improves every test set and can be repeated as better seed models become available.
  • The reported gains suggest that just adding more untranscribed audio to the same pipeline would lower WER further.

Reading between the lines

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

  • Beyond the paper: the same combination of continued pre-training, BPE units, language-model rescoring, and pseudo-labeling should transfer to other low-resource languages with comparable broadcast archives, although the compute for continued pre-training and the quality of the seed model are practical preconditions the paper does not quantify.
  • Beyond the paper: the final hybrid system could be run over much larger archives to produce pseudo-labels, and those labels could then be used to train a stronger end-to-end model, a direction the paper's future-work section points to but leaves untested.
  • Beyond the paper: the high WER on the BBC test set (17.7%) is concentrated in children's programmes with music and overlapping voices, so the recipe's next bottleneck is acoustically difficult conditions rather than language modelling or data quantity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper describes a practitioner-oriented recipe for building an ASR system for Scottish Gaelic, a low-resource language. The central empirical claim is that a carefully tuned hybrid HMM-DNN system using self-supervised features (XLS-R with continued pre-training), BPE acoustic units, subword n-gram and RNN language models, noise augmentation, active learning, and semi-supervised training on 184 hours of untranscribed news audio outperforms a fine-tuned Whisper-Turbo baseline by 32% relative WER and the previous Tarsgrìobhadair system by 54%. The paper provides a detailed ablation chain over feature extractors, acoustic units, language models, and data augmentation strategies, with results on four Gaelic test sets.

Significance. If the reported comparison is valid, the paper makes a valuable practical contribution: it demonstrates a concrete, reproducible recipe in which traditional hybrid ASR components combined with modern self-supervised models beat a widely used end-to-end fine-tuning approach in a low-resource setting. The strengths include the systematic ablations, the use of publicly available SSL models and toolkits, the comparison against an existing deployed system, and the use of real broadcast data with professional transcription. However, the headline claim rests on the integrity of the evaluation protocol, especially the separation between the semi-supervised training pool and the News test set. The paper also compares against a LoRA fine-tuned Whisper model rather than a fully fine-tuned one, and it reports a validation set as part of the test average. These issues need to be resolved before the main conclusion can be accepted.

major comments (4)
  1. [Section 4.6 / Table 1 / Section 3] The News test set is defined in Section 3 as seven episodes broadcast from 16 to 22 October 2023, but the paper never states that these episodes were excluded from the 184 hours of untranscribed An Là data used for semi-supervised training in Section 4.6. Table 1 labels this resource as 'News (untranscribed)', and Section 4.6 only says '300 hours of An Là data' with no date range. If any of the seven test episodes fall within the semi-supervised training pool, the reported News WER of 10.4%, the average WER of 12.8%, and the headline 32% relative improvement are invalid for that part of the evaluation. Please state explicitly the date range and programme identifiers of all An Là data used for training, and if overlap exists, rerun the evaluation on held-out episodes or otherwise demonstrate that the improvement survives uncontaminated evaluation.
  2. [Section 4.1 / Section 5] The Whisper baseline is only fine-tuned with LoRA on frozen 8-bit weights (ranks 64 to 512), not with full fine-tuning. The paper itself acknowledges in Section 5 that 'we might achieve even better results when fine-tuning the whole Whisper-Turbo model.' The abstract's phrase 'our best fine-tuned Whisper model' is therefore potentially misleading. Either run a full fine-tuning experiment to establish the true Whisper baseline, or scope the claim explicitly to 'our best LoRA fine-tuned Whisper-Turbo model' throughout the abstract, results, and conclusion. As written, the 32% relative improvement may understate the gap or overstate the advantage over the best possible fine-tuned Whisper.
  3. [Section 5 / Table 3 / Section 3] The CLTW test set (0.9 hours) is described in Section 3 as validation data: 'Due to its small size, we decided to use this dataset as a validation set during our experiments.' Nevertheless, Table 3 reports CLTW test as one of the four evaluation sets and includes it in the average WER used for the headline comparison. Using the same data for model selection and final reporting overstates generalization. Please remove CLTW from the reported test average or present it separately as a validation set, and recalculate the relative improvements without it. At minimum, the paper should clarify how CLTW was used during development and why it remains in the reported test suite.
  4. [Sections 3 and 5] No confidence intervals, bootstrap resampling, or significance tests are reported for any WER differences. Several test sets are very small (CLTW 0.9h, News 2.6h, PC 1.7h), and WER estimates on such sets can have wide variance. The 32% relative improvement is large, but the paper should accompany the key comparisons with uncertainty estimates (e.g., bootstrap confidence intervals) to confirm that the advantage over the Whisper baseline is not within noise. This is particularly important for the per-test-set numbers such as the News WER, which is central to the contamination concern.
minor comments (5)
  1. [Section 3] The word 'broadcasted' should be 'broadcast' in the description of the News test set, and the text has a LaTeX spacing artifact 'An L `a' that should render as 'An Là'.
  2. [Section 4.4] The BPE inventory sizes are listed as {500,1000,2000,5000}, and the choice of 1000 is justified later; consider stating the selection criterion (lowest average WER) in the same paragraph to avoid a brief ambiguity.
  3. [Section 5] The sentence 'Furthermore, we make additional use of all the untranscribed data, we decided to continue pre-training...' contains a comma splice; split it into two sentences for clarity.
  4. [References] References [2] and [3] have missing spaces in author names ('B. Shiet al.' and 'V . Sanhet al.') due to LaTeX formatting; these should be corrected.
  5. [Section 4.6 / Table 1] The text in Section 4.6 refers to 'An Là data' while Table 1 labels the same resource as 'News (untranscribed)'; please use consistent terminology, especially because the News test set is also called 'News'.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial evaluation circularity: CLTW is used as the validation set for model selection and then reported as a test set; no derivation-chain circularity elsewhere.

  1. fitted input called prediction [Section 3 (Data) and Section 5 / Table 3 (Results)]
    "Due to its small size, we decided to use this dataset as a validation set during our experiments. ... We used four datasets for evaluating our model: see Table 2."

    The same 0.9 h CLTW set is used to select all modelling choices reported in Section 4 (SSL feature extractor, BPE inventory 1000, RNN-LM size 2048, noise augmentation, extra data) by comparing average WER over the four evaluation sets, then Table 3 reports a 'CLTW test' WER for the final model and folds it into the 12.8% average that yields the headline 32% relative improvement over Whisper. The reported CLTW number is a component of the selection objective, not an independent prediction; it is fitted input labeled as test. The comparison on the PC heldout set (and on News/BBC if uncontaminated) remains independent, so the circularity is partial.

full rationale

The paper's central claim is an empirical comparison against external baselines (Whisper, Tar-sgrìobhadair), not a derivation from first principles; no equation-level circularity is present. Authors' prior work ([16], [18], [23], [24]) is used for data, baselines, and background techniques, and none of it is invoked as an unverified uniqueness or ansatz justification, so self-citation is not load-bearing. The one concrete reduction-by-construction issue is the CLTW test set doubling as the validation set, which makes the CLTW portion of the reported WER and of the average non-independent. Separately, the paper never states that the 184 h An Là semi-supervised pool excludes the seven News test episodes (16–22 Oct 2023) or the 18 Oct BBC newscast; if those episodes are in the pool, News/BBC results are contaminated, but this would be an evaluation-validity problem rather than a circular-reasoning step and cannot be confirmed from the text.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities or fitted physical constants. Its free parameters are engineering hyperparameters tuned on the validation set; they are many, but they are standard in ASR practice. The load-bearing unstated premise is data disjointness between the semi-supervised pool and the News test set.

free parameters (5)
  • BPE inventory size for acoustic units = 1000 tokens
    Chosen from {500, 1000, 2000, 5000} by validation WER; the best models with 1000 and 2000 both scored 16.0% average WER, and 1000 was selected.
  • LoRA rank for Whisper-Turbo baseline = 512
    Selected from {64, 128, 256, 512}; higher rank gave lower WER (22.0% average), and the paper notes full fine-tuning might do even better, so the baseline may understate E2E capability.
  • RNN-LM size = embedding 2048, LSTM cells 512
    Largest RNN-LM gave the best rescoring result (15.0% average WER); chosen without investigating larger sizes.
  • Continued pre-training iterations of XLS-R = 100k final, 40k intermediate
    40k steps used for initial comparison; final model continues pre-training on all data for 100k iterations; number chosen by compute budget, not by a stopping criterion.
  • SSL feature layer = 18th layer
    Section 4.3 says features were extracted from the 18th layer of both SSL models, with no ablation over layers; this choice affects all subsequent results.
assumptions (4)
  • domain assumption XLS-R 300M hidden representations are a suitable feature representation for Gaelic after continued self-supervised pre-training.
    Section 4.3 replaces MFCCs with the 18th layer of XLS-R/XEUS without theoretical justification; the assumption is validated only by the resulting WER.
  • domain assumption BPE subword tokens can serve as acoustic units in an LF-MMI TDNN-F system with bi-phones and tree clustering.
    Section 4.4 assumes that context-dependent BPE units behave like graphemes for acoustic modeling; no formal argument is given, only empirical results.
  • ad hoc to paper The 184 hours of untranscribed An Là news audio used for semi-supervised training does not overlap with the News test episodes.
    Section 3 (Tables 1 and 2) and Section 4.6. The paper states a holdout for PC test episodes but never states the same for the News test set, which is drawn from the same programme period.
  • domain assumption Semi-supervised pseudo-labels from the seed model are accurate enough to improve the acoustic model.
    Section 4.6 relies on the standard self-training assumption; supported by prior work [14, 15] but not directly verified for this data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Practitioner's Guide to Building ASR Models for Low-Resource Languages: A Case Study on Scottish Gaelic." pith.science (2026). https://pith.science/paper/VR3PRW64

@misc{pith2026250604915,
  author       = {Pith},
  title        = {Pith review of: A Practitioner's Guide to Building ASR Models for Low-Resource Languages: A Case Study on Scottish Gaelic},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VR3PRW64}},
  note         = {Machine review of arXiv:2506.04915}
}
read the original abstract

An effective approach to the development of ASR systems for low-resource languages is to fine-tune an existing multilingual end-to-end model. When the original model has been trained on large quantities of data from many languages, fine-tuning can be effective with limited training data, even when the language in question was not present in the original training data. The fine-tuning approach has been encouraged by the availability of public-domain E2E models and is widely believed to lead to state-of-the-art results. This paper, however, challenges that belief. We show that an approach combining hybrid HMMs with self-supervised models can yield substantially better performance with limited training data. This combination allows better utilisation of all available speech and text data through continued self-supervised pre-training and semi-supervised training. We benchmark our approach on Scottish Gaelic, achieving WER reductions of 32% relative over our best fine-tuned Whisper model.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 43 canonical work pages

  1. [1]

    For example, OpenAI’s Whisper [1] is a state-of-the-art mul- tilingual model that supports around 100 languages

    Introduction Automatic speech recognition (ASR) has been democratised by several public-domain multi-lingual end-to-end (E2E) models. For example, OpenAI’s Whisper [1] is a state-of-the-art mul- tilingual model that supports around 100 languages. Another example, Meta’s MMS [2], supports more than 1000 languages. However, this is only a fraction of all 70...

    arXiv 2025

  2. [2]

    Hybrid models [25] can effectively combine acoustic models trained on transcribed data with language mod- els trained on all available text

    Method To train the best possible Gaelic ASR system with a limited amount of manually transcribed data, it is important to leverage all available text corpora, which might be more accessible than transcribed speech. Hybrid models [25] can effectively combine acoustic models trained on transcribed data with language mod- els trained on all available text. ...

    work page

  3. [3]

    CLTW train

    Data We trained our models on several datasets: see Table 1. We used the same training data presented in [24] at the Celtic Lan- guage Technology Workshop, which we label “CLTW train”. This dataset consists of 103 hours of teaching videos, tradi- tional narratives, and audio books, which were automatically aligned with normalised transcripts. Upon analysi...

    work page 2023

  4. [4]

    Whisper Fine-Tuning We fine-tuned Whisper [1] with the HuggingFace Transformers library [3] as a baseline approach for training Gaelic ASR

    Experiments 4.1. Whisper Fine-Tuning We fine-tuned Whisper [1] with the HuggingFace Transformers library [3] as a baseline approach for training Gaelic ASR. Due to the limited amounts of transcribed training data (288 hours) and computation constraints, we decided to fine-tune Whisper- Turbo [1] with LoRA [29] using 8-bit quantisation of the frozen weight...

    work page 2000

  5. [5]

    The Tar-sgr`ıobhadair API achieved an average WER of 28.0%

    Results We used the Tar-sgr`ıobhadair API and Whisper as baselines for our experiments. The Tar-sgr`ıobhadair API achieved an average WER of 28.0%. Looking at fine-tuned Whisper-Turbo mod- els, we see that the performance improves with the number of fine-tuned parameters. The best average WER of 22.0% was achieved with a LoRA rank 512. This suggests that ...

    work page

  6. [6]

    Our model outperformed the previous best model deployed in the Tar-sgr`ıobhadair API by 54%

    Conclusions In this paper we showed that optimized hybrid models can achieve very good results on low-resource languages such as Scottish Gaelic. Our model outperformed the previous best model deployed in the Tar-sgr`ıobhadair API by 54%. Further- more, it outperformed our best fine-tuned Whisper-Turbo model by 32% relative. To get the best possible perfo...

    work page

  7. [7]

    We thank BBC Alba for providing the data and Cailean Gordan, Al- ison Diack and Fearchar MacIllFhinnein for transcribing train- ing and testing data

    Acknowledgements This work was supported by the Scottish Government (Grant name: ‘Ecosystem for Interactive Speech Technologies’). We thank BBC Alba for providing the data and Cailean Gordan, Al- ison Diack and Fearchar MacIllFhinnein for transcribing train- ing and testing data

    work page

  8. [8]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inICML, 2023

    2023

Show all 47 references
  1. [9]

    Scaling speech technology to 1,000+ languages,

    V . Pratap, A. Tjandra, B. Shiet al., “Scaling speech technology to 1,000+ languages,”JMLR, vol. 25, no. 97, pp. 1–52, 2024

    2024

  2. [10]

    Transformers: State-of-the-art natural language processing,

    T. Wolf, L. Debut, V . Sanhet al., “Transformers: State-of-the-art natural language processing,”EMNLP, 2020

    2020

  3. [11]

    Semi-Orthogonal Low-Rank Matrix Factor- ization for Deep Neural Networks,

    D. Povey, G. Cheng, Y . Wang, K. Li, H. Xu, M. Yarmohammadi, and S. Khudanpur, “Semi-Orthogonal Low-Rank Matrix Factor- ization for Deep Neural Networks,” inInterspeech, 2018

    2018

  4. [12]

    Principled comparisons for end-to-end speech recognition: Attention vs hybrid at the 1000- hour scale,

    A. Rouhe, T. Gr ´osz, and M. Kurimo, “Principled comparisons for end-to-end speech recognition: Attention vs hybrid at the 1000- hour scale,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 32, pp. 623–638, 2023

    2023

  5. [13]

    Madlad-400: A multilin- gual and document-level large audited dataset,

    S. Kudugunta, I. Caswell, B. Zhang, X. Garcia, D. Xin, A. Kusu- pati, R. Stella, A. Bapna, and O. Firat, “Madlad-400: A multilin- gual and document-level large audited dataset,”NeurIPS, 2024

    2024

  6. [14]

    The fineweb datasets: Decanting the web for the finest text data at scale,

    G. Penedo, H. Kydl ´ıˇcek, L. Ben allal, A. Lozhkov, M. Mitchell, C. A. Raffel, L. V on Werra, and T. Wolf, “The fineweb datasets: Decanting the web for the finest text data at scale,” inNeurIPS, 2024

    2024

  7. [15]

    Recurrent neural network based language model

    T. Mikolov, M. Karafi ´at, L. Burget, J. Cernock `y, and S. Khudan- pur, “Recurrent neural network based language model.” inInter- speech, 2010

    2010

  8. [16]

    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

  9. [17]

    XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale,

    A. Babu, C. Wang, A. Tjandraet al., “XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale,” inInter- speech, 2022

    2022

  10. [18]

    Towards robust speech representation learning for thousands of languages,

    W. Chen, W. Zhang, Y . Peng, X. Li, J. Tian, J. Shi, X. Chang, S. Maiti, K. Livescu, and S. Watanabe, “Towards robust speech representation learning for thousands of languages,” inEMNLP, 2024

    2024

  11. [19]

    The language-independent bottleneck features,

    K. Vesel ´y, M. Karafi´at, F. Gr´ezl, M. Janda, and E. Egorova, “The language-independent bottleneck features,” inSLT, 2012

    2012

  12. [20]

    CTRL: Continual Representation Learning to Transfer In- formation of Pre-trained for W A V2VEC 2.0,

    J.-H. Lee, C.-W. Lee, J.-S. Choi, J.-H. Chang, W. K. Seong, and J. Lee, “CTRL: Continual Representation Learning to Transfer In- formation of Pre-trained for W A V2VEC 2.0,”Interspeech, 2022

    2022

  13. [21]

    Lightly supervised and unsupervised acoustic model training,

    L. Lamel, J.-L. Gauvain, and G. Adda, “Lightly supervised and unsupervised acoustic model training,”Computer Speech and Language, vol. 16, no. 1, pp. 115–229, 2002

    2002

  14. [22]

    Semi- supervised training of acoustic models using lattice-free MMI,

    V . Manohar, H. Hadian, D. Povey, and S. Khudanpur, “Semi- supervised training of acoustic models using lattice-free MMI,” inICASSP, 2018

    2018

  15. [23]

    On the learning dynamics of semi-supervised training for ASR,

    E. Wallington, B. Kershenbaum, O. Klejch, and P. Bell, “On the learning dynamics of semi-supervised training for ASR,” inInter- speech, 2021

    2021

  16. [24]

    Self-training and pre- training are complementary for speech recognition,

    Q. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and M. Auli, “Self-training and pre- training are complementary for speech recognition,” inICASSP, 2021

    2021

  17. [25]

    Compar- ing self-supervised pre-training and semi-supervised training for speech recognition in languages with weak language models,

    L.-M. Lam-Yee-Mui, L. O. Yang, and O. Klejch, “Compar- ing self-supervised pre-training and semi-supervised training for speech recognition in languages with weak language models,” in Interspeech, 2023

    2023

  18. [26]

    Improved noisy student training for automatic speech recognition,

    D. S. Park, Y . Zhang, Y . Jia, W. Han, C.-C. Chiu, B. Li, Y . Wu, and Q. V . Le, “Improved noisy student training for automatic speech recognition,” inInterspeech, 2020

    2020

  19. [27]

    Cross- lingual knowledge transfer and iterative pseudo-labeling for low- resource speech recognition with transducers,

    J. Silovsky, L. Deng, A. Argueta, T. Arvizo, R. Hsiao, S. Kuznietsov, Y .-C. Lin, X. Xiao, and Y . Zhang, “Cross- lingual knowledge transfer and iterative pseudo-labeling for low- resource speech recognition with transducers,”arXiv preprint arXiv:2305.13652, 2023

    2023 arXiv

  20. [28]

    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

  21. [29]

    Scotland’s census 2022,

    National Records of Scotland, “Scotland’s census 2022,” https: //www.scotlandscensus.gov.uk, 2022, accessed: 2024-08-14

    2022

  22. [30]

    Grapheme and multilin- gual posterior features for under-resourced speech recognition: a study on Scottish Gaelic,

    R. Rasipuram, P. Bell, and M. M. Doss, “Grapheme and multilin- gual posterior features for under-resourced speech recognition: a study on Scottish Gaelic,” inICASSP, 2013

    2013

  23. [31]

    Developing auto- matic speech recognition for Scottish Gaelic,

    L. Evans, W. Lamb, M. Sinclair, and B. Alex, “Developing auto- matic speech recognition for Scottish Gaelic,” inCeltic Language Technology Workshop, 2022

    2022

  24. [32]

    H. A. Bourlard and N. Morgan,Connectionist speech recognition: a hybrid approach. Springer Science & Business Media, 2012, vol. 247

    2012

  25. [33]

    Advances in subword- based HMM-DNN speech recognition across languages,

    P. Smit, S. Virpioja, and M. Kurimo, “Advances in subword- based HMM-DNN speech recognition across languages,”Com- puter Speech & Language, vol. 66, p. 101158, 2021

    2021

  26. [34]

    Active and semi- supervised learning in ASR: Benefits on the acoustic and language models,

    T. Drugman, J. Pylkk ¨onen, and R. Kneser, “Active and semi- supervised learning in ASR: Benefits on the acoustic and language models,” inInterspeech, 2016

    2016

  27. [35]

    The MGB challenge: Evalu- ating multi-genre broadcast media recognition,

    P. Bell, M. J. Gales, T. Hainet al., “The MGB challenge: Evalu- ating multi-genre broadcast media recognition,” inASRU, 2015

    2015

  28. [36]

    LoRA: Low-Rank Adaptation of Large Language Models,

    E. J. Hu, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “LoRA: Low-Rank Adaptation of Large Language Models,” inICLR, 2022

    2022

  29. [37]

    Purely sequence-trained neu- ral networks for ASR based on lattice-free MMI,

    D. Povey, V . Peddinti, D. Galvez, P. Ghahremani, V . Manohar, X. Na, Y . Wang, and S. Khudanpur, “Purely sequence-trained neu- ral networks for ASR based on lattice-free MMI,” inInterspeech, 2016

    2016

  30. [38]

    SRILM - an extensible language modeling toolkit

    A. Stolckeet al., “SRILM - an extensible language modeling toolkit.” inInterspeech, 2002

    2002

  31. [39]

    fairseq: A fast, extensible toolkit for sequence modeling,

    M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grang- ier, and M. Auli, “fairseq: A fast, extensible toolkit for sequence modeling,” inNAACL-HLT: Demonstrations, 2019

    2019

  32. [40]

    From senones to chenones: Tied context-dependent graphemes for hybrid speech recognition,

    D. Le, X. Zhang, W. Zheng, C. F ¨ugen, G. Zweig, and M. L. Seltzer, “From senones to chenones: Tied context-dependent graphemes for hybrid speech recognition,” inASRU, 2019

    2019

  33. [41]

    Faster, simpler and more accurate hybrid ASR systems using wordpieces,

    F. Zhang, Y . Wang, X. Zhang, C. Liu, Y . Saraf, and G. Zweig, “Faster, simpler and more accurate hybrid ASR systems using wordpieces,” inInterspeech, 2020

    2020

  34. [42]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” inICASSP, 2017

    2017

  35. [43]

    Unsupervised language model adap- tation,

    M. Bacchiani and B. Roark, “Unsupervised language model adap- tation,” inICASSP, 2003

    2003

  36. [44]

    Less is More: Accu- rate Speech Recognition & Translation without Web-Scale Data,

    K. C. Puvvada, P. Zelasko, H. Huanget al., “Less is More: Accu- rate Speech Recognition & Translation without Web-Scale Data,” inInterspeech, 2024

    2024

  37. [45]

    Latxa: An open lan- guage model and evaluation suite for Basque,

    J. Etxaniz, O. Sainz, N. Miguel, I. Aldabe, G. Rigau, E. Agirre, A. Ormazabal, M. Artetxe, and A. Soroa, “Latxa: An open lan- guage model and evaluation suite for Basque,” inACL, 2024

    2024

  38. [46]

    Hyporadise: An open baseline for generative speech recognition with large language models,

    C. Chen, Y . Hu, C.-H. H. Yang, S. M. Siniscalchi, P.-Y . Chen, and E.-S. Chng, “Hyporadise: An open baseline for generative speech recognition with large language models,”NeurIPS, 2024

    2024

  39. [47]

    LLaMA-omni: Seamless speech interaction with large language models,

    Q. Fang, S. Guo, Y . Zhou, Z. Ma, S. Zhang, and Y . Feng, “LLaMA-omni: Seamless speech interaction with large language models,” inICLR, 2025

    2025

Pith tools

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