Evaluating Standard and Dialectal Frisian ASR: Multilingual Fine-tuning and Language Identification for Improved Low-resource Performance

Jelske Dijkstra; Martijn Wieling; Matt Coler; Reihaneh Amooie; Wietse de Vries; Yun Hao

REVIEW 3 major objections 6 minor 17 references

Evaluating Standard and Dialectal Frisian ASR: Multilingual Fine-tuning and Language Identification for Improved Low-resource Performance

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Frisian ASR improves when Dutch and German are added to fine-tuning data.

desk verdict A useful first look at Frisian dialect ASR with a new corpus, but the claimed multilingual gain is confounded with training-data size and the paper needs an equal-data control and a clearer LID protocol. read the letter →

arxiv 2502.04883 v1 pith:2KMYH3DJ submitted 2025-02-07 cs.CL cs.LGcs.SDeess.AS

Reihaneh Amooie , Wietse de Vries , Yun Hao , Jelske Dijkstra , Matt Coler , Martijn Wieling This is my paper · ORCID

classification cs.CLcs.LGcs.SDeess.AS

keywords automaticspeechrecognitionFrisianlow-resourcelanguagesmultilingualfine-tuninglanguageidentificationdialectalself-supervisedlearningworderrorrate

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

The paper tries to establish that a low-resource ASR system for Frisian can be improved by fine-tuning a self-supervised multilingual model on closely related languages and by giving the model a language-identification hint. On standard read Frisian, word error rate drops from 14.2 percent (Frisian-only) to 13.1 percent when Dutch and German are added, while adding English hurts. On dialectal speech, all models perform far worse, and the paper argues the gap reflects genuine dialect variation rather than merely difficult audio. It also claims that how dialect data are elicited matters: speakers who translate from Dutch texts produce more regionally marked speech than speakers who read Standard Frisian aloud, so standard-only benchmarks may overstate real-world ASR accuracy. A sympathetic reading takes these results as evidence that multilingual fine-tuning plus language identification is a useful recipe for minority languages with dialectal variation.

What carries the argument

The mechanism is staged multilingual fine-tuning of the self-supervised XLS-R 1B model with connectionist temporal classification loss, updating the encoder layers while freezing the feature extractor. Language similarity ordering is supplied by lexical-phonetic distances derived from ASJP data, so the paper adds Dutch, then German, then English, downsampling each added language to match the 3,921 Frisian sentences to prevent bias toward higher-resource languages. Language identification is implemented as a token prepended to the target transcription, which allows the model to condition its decoding on the predicted language without a separate classifier. Dialectal evaluation comes from the SPRAAKLAB mobile-lab recordings, where sixteen elderly male speakers produced the same story content in Clay, Wood, or South Frisian under two elicitation conditions: reading Standard Frisian text versus translating Dutch text into their own dialect.

What would settle it

Re-run the SPRAAKLAB evaluation with the same speakers producing both elicitation conditions, or add an independent transcription check; if the Dutch-elicited versus Frisian-elicited WER gap shrinks to zero when speaker identity is held fixed, the claimed elicitation effect collapses. A simpler check is to measure inter-annotator agreement on the dialectal transcriptions, because if disagreement is comparable to the one-to-two percent WER differences being interpreted, the moderation claim is not distinguishable from transcription noise.

Watch

Extended reading notes

Core claim

The central discovery is a measured word-error-rate comparison across staged fine-tuning configurations. Starting from an XLS-R 1B model fine-tuned on Frisian only, which reaches 14.2 percent test WER on standard Frisian, adding Dutch lowers the error to 13.6 percent and adding German further lowers it to 13.1 percent; adding English reverses the gain to 13.4 percent. An auxiliary language-identification token prepended to each target transcription gives a marginal further improvement on standard data and a clearer reduction of about one percent on dialectal SPRAAKLAB data. The same models show 24 to 27 percent WER on dialectal speech, and the paper attributes this residual gap to dialectal variation plus a systematic elicitation effect: Dutch-prompted translations yield more dialectal output and are harder to recognize than Standard Frisian-prompted output. The paper therefore claims that dialectal evaluation should be a standard part of low-resource ASR assessment, not an optional extra.

Load-bearing premise

The dialectal conclusions stand on the assumption that the two SPRAAKLAB speaker groups differ only in prompt language and that the manual transcriptions of dialectal utterances are accurate, since no inter-annotator agreement or speaker overlap is reported.

Editorial extensions

If this is right

Adding one or two closely related languages to fine-tuning data can lower WER for a low-resource target language, while adding a more distant language can hurt performance.
Language-identification tokens provide a larger benefit when the test domain is harder, such as dialectal speech, than on standard read speech.
Standard Frisian test sets underestimate field performance, since dialectal speech shows roughly ten percentage points higher word error than standard speech.
Evaluation stimuli presented in the standard language produce less dialectal output and therefore flatter results, so translation-elicited dialect data offers a more realistic challenge for minority-language ASR.
The ordering of added languages by linguistic distance appears to matter, because English, the least similar language tried, degrades the gain from Dutch and German.

Reading between the lines

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

The paper does not test whether the similarity ordering itself is the active ingredient; an extension could add other related West Germanic varieties, such as Low Saxon, and check whether gains track lexical-phonetic distance monotonically.
The paper leaves open whether LID tokens work by giving the decoder a prior over the output language or by regularizing the encoder; ablating the token position or applying it at the encoder output could separate these mechanisms.
The persistently high dialectal WER suggests that more multilingual data alone will not close the dialect gap, and that dialect-aware acoustic models, dialect lexicons, or additional dialectal training data are likely needed; this goes beyond what the paper demonstrates.
If the elicitation effect is real, ASR builders for other dialectally diverse low-resource languages should collect evaluation data using translation prompts, not only read-standard prompts, to avoid optimistic benchmarks.

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.

Desk Editor's Note

read the letter

Worth a read if you work on low-resource ASR. The genuinely new thing is the SPRAAKLAB dialect corpus and the first reported evaluation of multilingual fine-tuning plus LID on Frisian's three dialects. Two elicitation modes (translate from Dutch vs. read Standard Frisian) is a nice design idea, and the data collection is reported carefully: sixteen elderly men, balanced across regions, 103 sentences each. The ASJP-based ordering of language addition and the downsampling to equal sentence counts are sensible methodological choices.

That said, the central claim is shakier than the abstract suggests. The experimental design varies language mixture and total training-data volume at the same time: the baseline is 5.5h of Frisian, the best system is 16h. Table II shows even the four-language model with English beats the 5.5h baseline, and English is the most distant language. A pure data-volume effect predicts exactly that pattern, so the attribution to multilingual transfer is not identified. There are no confidence intervals or seed variation, so the 1.1-point WER gap could be noise.

The LID part also has an unresolved oracle problem. The paper says LID tokens are prepended to the target transcription, but it never says what happens at test time. If the correct language token is given to the decoder, the reported LID benefit on SPRAAKLAB is inflated. If it is predicted, the model's LID recall should be tied to actual usage; either way, readers need this spelled out.

The dialectal evaluation is under-powered: sixteen elderly men, no inter-annotator agreement on the manual transcriptions, no speaker overlap between the two elicitation groups, and no released data to check comparability. The moderating effect of elicitation rests on a small sample and is plausible but not proven.

Who is this for? Practitioners building ASR for small West Germanic varieties, and dialectologists who care about how evaluation data is elicited. The corpus alone justifies referee time. I would send it to peer review, but with a clear expectation: add an equal-data monolingual control (e.g., 16h of Frisian if available, or a subsampled comparison), report variance across seeds, clarify test-time LID conditioning, and release or at least describe the SPRAAKLAB transcriptions in more detail. The paper is honest and the writing is clear; it just needs the experiments to back the headline claim.

Referee Report

3 major / 6 minor

Summary. The paper describes experiments fine-tuning XLS-R 1B on Frisian and on combinations of Frisian with Dutch, German, and English from Common Voice 17.0, with and without language-identification (LID) tokens, and evaluates the resulting models on Standard Frisian (CV test) and on a newly collected dialectal Frisian corpus (SPRAAKLAB). The authors report that adding Dutch and German to the fine-tuning data reduces WER on Standard Frisian from 14.2% to 13.1%, that LID tokens give small gains on standard speech and about 1% WER gains on dialectal speech, that dialectal speech is much harder than standard speech, and that the elicitation language (prompts in Dutch vs. Frisian) is associated with differences in dialectal WER. The paper concludes that multilingual fine-tuning and auxiliary LID improve low-resource Frisian ASR and that evaluating only on standard-language test sets may underestimate real-world performance on dialectal speech.

Significance. If fully supported, the paper would offer a practically useful fine-tuning recipe for a low-resource West Germanic language and a cautionary result about dialectal evaluation. Strengths include the use of a public pre-trained SSL model, realistic fine-tuning data, a dedicated dialectal corpus, and reporting of LID recall on the test sets. However, the central comparisons do not isolate the effect of multilingual transfer from the effect of simply adding more training data, and the dialectal evaluation rests on a small corpus whose validity and comparability across elicitation conditions are not established. The paper is therefore more suggestive than demonstrative in its current form, but the claims are plausible and the underlying approach is worth testing more rigorously.

major comments (3)

[III-A.1 and Table II] The claimed multilingual benefit is confounded with total training-data volume. The baseline (fy) uses 5.5h/3,921 sentences, while the best multilingual model (fy-nl-de) uses 16.0h/11,763 sentences. The 14.2% to 13.1% WER improvement is therefore compatible with a pure data-volume effect. Table II strengthens this concern: fy-nl-de-en (22.0h) still beats the 5.5h baseline (13.4% vs. 14.2%) even though English is the most distant language and adding it is detrimental relative to fy-nl-de. To support the 'multilingual transfer' interpretation, the paper needs an equal-data control (e.g., a Frisian-only model trained on the same number of sentences, if more Frisian data is available, or a multilingual model downsampled to 5.5h), or Frisian-only learning curves that allow extrapolation to 11,763 and 15,684 sentences. Additionally, no repeated runs, confidence intervals, or significance tests are reported, so the 1.1-point WER gap may also lie within run-to-run variability. Without these controls, the abstract's causal statement that ASR performance 'can be improved by using multilingual fine-tuning data' is not supported.
[III-A.3 and Table IV] The dialectal evaluation corpus is small and its validity is insufficiently established for the 'elicitation moderates performance' claim. The SPRAAKLAB data contain 824 sentences per elicitation subset from 16 speakers, all men over 60, with 8 speakers per subset and no reported speaker overlap between the Dutch-prompt and Frisian-prompt conditions. The manual transcriptions are mentioned but no inter-annotator agreement or transcription quality measure is reported. The comparison between WERSF and WERD is therefore confounded with speaker identity and potentially with transcription noise. The finding that Dutch stimuli produce more dialectally variable speech (and somewhat higher WER) is one of the headline conclusions, but as presented it is not a controlled comparison. The claim should either be supported with additional validation (e.g., speaker-balanced comparisons, transcription checks) or softened.
[III-B.2 and Tables III-IV] The test-time use of the LID token is under-specified, which is necessary to interpret the LID benefit. The paper says the token is added 'at the beginning of each target transcription' during training, but does not state whether at inference the token is forced to the known language or predicted by the decoder. The reported LID recall values (e.g., 99.7 for fy-nl) suggest the token is predicted, but the evaluation scheme should be explicit. If the token is instead provided as an oracle at test time, the approximately 1% WER gain on dialectal data is an upper bound rather than a practical auxiliary-task benefit. This missing detail is required to reproduce the experiments and to judge whether the LID results support the paper's claims.

minor comments (6)

[III-A.3] Please state whether the SPRAAKLAB dialectal corpus will be made available; a data-availability statement is needed for reproducibility.
[III-B.2] Please specify the early-stopping criterion and how many independent runs were performed for each configuration; the paper appears to report single runs.
[III-B.2] Typo: 'V ocabulary size' should be 'Vocabulary size'.
[References] Several references have inconsistent author formatting (e.g., 'A Vaswani' and 'Alex Graves and Alex Graves'); please unify the reference style.
[III-A.1 and III-A.3] The paper uses both '16 KHz' and '16kHz'; please standardize the notation.
[Table IV] The meaning of 'LIDSF recall' and 'LIDD recall' should be defined precisely when the test set is always Frisian; clarify whether these are the recall of the expected Frisian token on each subset and why the model is not always forced to emit the Frisian token.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: all headline claims are evaluated on held-out Frisian and externally collected dialectal test data.

full rationale

The paper's central claims are empirical, not derivational. Standard Frisian WER is measured on the official Common Voice 17.0 test split, and dialectal WER is measured on SPRAAKLAB audio collected independently from the fine-tuning data; neither outcome is defined in terms of the training-set composition or the fitted model parameters. The comparison across fy, fy-nl, fy-nl-de, and fy-nl-de-en varies language mixture together with total fine-tuning data volume, which is a genuine experimental confound and a threat to the causal attribution of the gain to multilingual transfer, but it is not circularity: the reported WER values are not forced by construction, and the paper contains no equation or definition that makes the improvement true by definition. The authors' prior work is cited for methodological choices only ([13] for lexical-phonetic similarity ordering and [14] for the mobile recording lab), and neither citation carries the load of the empirical conclusion. The LID-token procedure is under-specified at inference time, so a possible oracle-token inflation of the LID benefit cannot be excluded, but that would be an evaluation artifact rather than a self-referential derivation. No step in the paper reduces a prediction to its own input, so the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper is empirical, so the ledger lists the domain assumptions the central claims depend on rather than mathematical axioms. Corpus labeling, transferability of XLS-R, language similarity ordering, and manual transcription validity are all assumed without independent verification in the paper.

free parameters (3)

Learning rate = 0.00008
Fixed by hand for all fine-tuning runs; no search or sensitivity analysis is reported, so absolute WER values may depend on this choice.
Batch size = 8 with 16 gradient accumulation steps
Chosen without sensitivity analysis; affects effective batch size and training dynamics.
Maximum training epochs = 50
Fixed stopping point; no early stopping or convergence analysis is described.

assumptions (5)

domain assumption Common Voice 17.0 split labels are correct and representative of Standard Frisian.
The baseline and evaluation assume the CV Frisian data is a valid sample of standard Frisian speech; the paper does not audit transcription quality. Invoked in Section III-A.1.
domain assumption XLS-R pre-trained representations transfer to Frisian and its dialects after fine-tuning.
The method assumes self-supervised multilingual pre-training provides useful features for a low-resource target language; this is the premise of the transfer approach. Invoked in Section III-B.1.
domain assumption ASJP lexical-phonetic distances correctly order language similarity for the incremental addition of Dutch, German, and English.
The order of language addition is based on ASJP distances cited from prior work; if this ordering is wrong, the interpretation of language proximity effects changes. Invoked in Section III-A.1.
domain assumption Manual transcriptions of SPRAAKLAB dialectal speech are accurate.
No inter-annotator agreement or validation is reported; labels are treated as ground truth for WER computation. Invoked in Section III-A.3.
domain assumption The two SPRAAKLAB subsets (Dutch vs Frisian prompts) are comparable despite being spoken by different speakers.
Each subset has different speakers (8 per condition); the comparison assumes speaker differences do not drive the observed WER difference. Invoked in Section III-A.3.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Evaluating Standard and Dialectal Frisian ASR: Multilingual Fine-tuning and Language Identification for Improved Low-resource Performance." pith.science (2026). https://pith.science/paper/2KMYH3DJ

@misc{pith2026250204883,
  author       = {Pith},
  title        = {Pith review of: Evaluating Standard and Dialectal Frisian ASR: Multilingual Fine-tuning and Language Identification for Improved Low-resource Performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KMYH3DJ}},
  note         = {Machine review of arXiv:2502.04883}
}

read the original abstract

Automatic Speech Recognition (ASR) performance for low-resource languages is still far behind that of higher-resource languages such as English, due to a lack of sufficient labeled data. State-of-the-art methods deploy self-supervised transfer learning where a model pre-trained on large amounts of data is fine-tuned using little labeled data in a target low-resource language. In this paper, we present and examine a method for fine-tuning an SSL-based model in order to improve the performance for Frisian and its regional dialects (Clay Frisian, Wood Frisian, and South Frisian). We show that Frisian ASR performance can be improved by using multilingual (Frisian, Dutch, English and German) fine-tuning data and an auxiliary language identification task. In addition, our findings show that performance on dialectal speech suffers substantially, and, importantly, that this effect is moderated by the elicitation approach used to collect the dialectal data. Our findings also particularly suggest that relying solely on standard language data for ASR evaluation may underestimate real-world performance, particularly in languages with substantial dialectal variation.

Figures

Figures reproduced from arXiv: 2502.04883 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 11 canonical work pages

[1]

Improving massively multilingual asr with auxiliary ctc objectives,

William Chen, Brian Yan, Jiatong Shi, Yifan Peng, Soumi Maiti, and Shinji Watanabe, “Improving massively multilingual asr with auxiliary ctc objectives,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

work page 2023
[2]

Hierarchical softmax for end-to-end low-resource multilin- gual speech recognition,

Qianying Liu, Zhuo Gong, Zhengdong Yang, Yuhang Yang, Sheng Li, Chenchen Ding, Nobuaki Minematsu, Hao Huang, Fei Cheng, Chenhui Chu, et al., “Hierarchical softmax for end-to-end low-resource multilin- gual speech recognition,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

work page 2023
[3]

A survey of multilingual models for automatic speech recognition,

Hemant Yadav and Sunayana Sitaram, “A survey of multilingual models for automatic speech recognition,” arXiv preprint arXiv:2202.12576 , 2022

arXiv 2022
[4]

Rapid language adaptation for multilingual e2e speech recognition using encoder prompting,

Yosuke Kashiwagi, Hayato Futami, Emiru Tsunoo, Siddhant Arora, and Shinji Watanabe, “Rapid language adaptation for multilingual e2e speech recognition using encoder prompting,” arXiv preprint arXiv:2406.12611, 2024

arXiv 2024
[5]

Mole: Mixture of language experts for multi-lingual automatic speech recognition,

Yoohwan Kwon and Soo-Whan Chung, “Mole: Mixture of language experts for multi-lingual automatic speech recognition,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

work page 2023
[6]

Leveraging language id in multilingual end-to-end speech recognition,

Austin Waters, Neeraj Gaur, Parisa Haghani, Pedro Moreno, and Zhongdi Qu, “Leveraging language id in multilingual end-to-end speech recognition,” in 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . IEEE, 2019, pp. 928–935

work page 2019
[7]

Streaming end-to-end multi- lingual speech recognition with joint language identification,

Chao Zhang, Bo Li, Tara Sainath, Trevor Strohman, Sepand Mavandadi, Shuo-Yiin Chang, and Parisa Haghani, “Streaming end-to-end multi- lingual speech recognition with joint language identification,” arXiv preprint arXiv:2209.06058, 2022

arXiv 2022
[8]

Speechmoe2: Mixture- of-experts model with improved routing,

Zhao You, Shulin Feng, Dan Su, and Dong Yu, “Speechmoe2: Mixture- of-experts model with improved routing,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7217–7221

work page 2022

Show all 17 references

[9]

Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

William Fedus, Barret Zoph, and Noam Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,” Journal of Machine Learning Research, vol. 23, no. 120, pp. 1–39, 2022

2022
[10]

Connectionist temporal classification,

Alex Graves and Alex Graves, “Connectionist temporal classification,” Supervised sequence labelling with recurrent neural networks , pp. 61– 93, 2012

2012
[11]

Common voice: A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020) , 2020, pp. 4211–4215

2020
[12]

Evaluating linguistic distance measures,

Søren Wichmann, Eric W Holman, Dik Bakker, and Cecil H Brown, “Evaluating linguistic distance measures,” Physica A: Statistical Me- chanics and its Applications , vol. 389, no. 17, pp. 3632–3639, 2010

2010
[13]

Adapting monolingual models: Data can be scarce when language similarity is high,

Wietse De Vries, Martijn Bartelds, Malvina Nissim, and Martijn Wiel- ing, “Adapting monolingual models: Data can be scarce when language similarity is high,” arXiv preprint arXiv:2105.02855 , 2021

2021 arXiv
[14]

Spraaklab: a mobile laboratory for collecting speech production data,

Martijn Wieling, Teja Rebernik, and Jidde Jacobi, “Spraaklab: a mobile laboratory for collecting speech production data,” in Proceedings of the 20th International Congress of Phonetic Sciences . Guarant International, 2023, pp. 2060–2064

2023
[15]

Dialect variation in and around frisia; classification and relationships,

Wilbert Heeringa, “Dialect variation in and around frisia; classification and relationships,” Us Wurk, vol. 54, no. 3-4, pp. 125–167, 2005

2005
[16]

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

Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick V on Platen, Yatharth Saraf, Juan Pino, et al., “Xls-r: Self-supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021

2021 arXiv
[17]

Attention is all you need,

A Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Improving massively multilingual asr with auxiliary ctc objectives,

William Chen, Brian Yan, Jiatong Shi, Yifan Peng, Soumi Maiti, and Shinji Watanabe, “Improving massively multilingual asr with auxiliary ctc objectives,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

work page 2023

[2] [2]

Hierarchical softmax for end-to-end low-resource multilin- gual speech recognition,

Qianying Liu, Zhuo Gong, Zhengdong Yang, Yuhang Yang, Sheng Li, Chenchen Ding, Nobuaki Minematsu, Hao Huang, Fei Cheng, Chenhui Chu, et al., “Hierarchical softmax for end-to-end low-resource multilin- gual speech recognition,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

work page 2023

[3] [3]

A survey of multilingual models for automatic speech recognition,

Hemant Yadav and Sunayana Sitaram, “A survey of multilingual models for automatic speech recognition,” arXiv preprint arXiv:2202.12576 , 2022

arXiv 2022

[4] [4]

Rapid language adaptation for multilingual e2e speech recognition using encoder prompting,

Yosuke Kashiwagi, Hayato Futami, Emiru Tsunoo, Siddhant Arora, and Shinji Watanabe, “Rapid language adaptation for multilingual e2e speech recognition using encoder prompting,” arXiv preprint arXiv:2406.12611, 2024

arXiv 2024

[5] [5]

Mole: Mixture of language experts for multi-lingual automatic speech recognition,

Yoohwan Kwon and Soo-Whan Chung, “Mole: Mixture of language experts for multi-lingual automatic speech recognition,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

work page 2023

[6] [6]

Leveraging language id in multilingual end-to-end speech recognition,

Austin Waters, Neeraj Gaur, Parisa Haghani, Pedro Moreno, and Zhongdi Qu, “Leveraging language id in multilingual end-to-end speech recognition,” in 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . IEEE, 2019, pp. 928–935

work page 2019

[7] [7]

Streaming end-to-end multi- lingual speech recognition with joint language identification,

Chao Zhang, Bo Li, Tara Sainath, Trevor Strohman, Sepand Mavandadi, Shuo-Yiin Chang, and Parisa Haghani, “Streaming end-to-end multi- lingual speech recognition with joint language identification,” arXiv preprint arXiv:2209.06058, 2022

arXiv 2022

[8] [8]

Speechmoe2: Mixture- of-experts model with improved routing,

Zhao You, Shulin Feng, Dan Su, and Dong Yu, “Speechmoe2: Mixture- of-experts model with improved routing,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7217–7221

work page 2022