Continual Adaptation for Pacific Indigenous Speech Recognition

Aso Mahmudi; Eliathamby Ambikairajah; Eun-Jung Holden; Nick Thieberger; Ting Dang; Yang Xiao

REVIEW 4 major objections 6 minor 39 references

Continual Adaptation for Pacific Indigenous Speech Recognition

T0 review · 4 major / 6 minor · reviewed 2026-07-15 · grok-4.5

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

Pith's one-line read Adapting speech foundation models to distant Pacific Indigenous languages induces severe representational drift and creates a hard plasticity-stability dilemma.

desk verdict Clean empirical demonstration that Whisper adaptation to real Pacific languages forces a hard plasticity-stability bind, even if the drift-to-forgetting causal link stays correlational. read the letter →

arxiv 2603.06310 v2 pith:UUR6K2MD submitted 2026-03-06 eess.AS cs.CLcs.SD

Yang Xiao , Aso Mahmudi , Nick Thieberger , Eliathamby Ambikairajah , Eun-Jung Holden , Ting Dang This is my paper · ORCID

classification eess.AScs.CLcs.SD

keywords speechrecognitionPacificlanguagescontinuallearningrepresentationaldriftcatastrophicforgettingLoRAlow-resourceASRIndigenous

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

Speech foundation models fail to transfer cleanly to real low-resource Pacific languages because those languages sit far from the high-resource data used in pretraining. Fine-tuning therefore forces large internal reorganisation of the model's acoustic and linguistic features rather than simple refinement. The paper measures this reorganisation as layer-wise representational drift and shows that it produces a strict trade-off: methods that quickly acquire a new Pacific language (especially LoRA) erase earlier languages, while full fine-tuning retains more prior knowledge but underperforms on the new target under extreme data scarcity. Sequential experiments on Nafsan then Lelepa confirm that neither standard approach solves the dilemma. The result matters because it demonstrates that today's adaptation toolkit is structurally inadequate for underrepresented Indigenous languages and that speech technology for Pacific communities will require purpose-built continual-learning methods.

What carries the argument

Layer-wise representational drift quantified by cosine distance between pre- and post-adaptation hidden states of every encoder and decoder layer, evaluated on a fixed multi-language set and linked to sequential continual-learning error rates that expose catastrophic forgetting.

What would settle it

If high measured cosine drift occurred without any rise in error rates on previously learned languages, or if low-drift adaptations still produced severe forgetting on English and earlier Pacific languages, the claimed causal link between drift and the plasticity-stability dilemma would be disproved.

Watch

Extended reading notes

Core claim

Adapting speech foundation models to linguistically distant Pacific Indigenous languages induces severe internal representational drift. As a direct consequence the models confront a strict plasticity-stability dilemma: LoRA and related parameter-efficient updates achieve good initial accuracy on the new target yet suffer catastrophic forgetting of prior languages during sequential learning, while full fine-tuning better preserves earlier knowledge at the cost of weaker acquisition of the new language.

Load-bearing premise

The paper treats cosine distance between pre- and post-adaptation hidden states, after per-language min-max normalisation, as a faithful measure of the structural reorganisation that causes forgetting.

Editorial extensions

If this is right

Neither full fine-tuning nor LoRA (nor DoRA/O-LoRA) can simultaneously acquire a new distant Pacific language and retain earlier ones.
Adaptation success scales with data volume and linguistic proximity; English-related creoles adapt far more readily than isolated Austronesian languages.
Updating only the encoder or only the decoder fails to balance target accuracy against source-language retention.
Speech systems intended for Pacific communities will need continual-learning methods that explicitly protect historical representations.
Future architectures must handle both extreme data imbalance and deep acoustic restructuring.

Reading between the lines

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

Comparable drift-and-forgetting patterns are likely for other typologically distant low-resource families outside the Pacific.
Regularisation techniques successful in text continual learning may require acoustic-specific redesign before they protect speech foundation models.
Modular, language-isolated adapters could reduce interference without sacrificing the plasticity needed for extreme low-resource targets.
Multilingual ASR benchmarks should report representational drift and sequential forgetting alongside final error rates.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

The one thing worth knowing is that this paper actually measures what happens when you push Whisper-Small onto three genuine low-resource Pacific languages (Bislama, Nafsan, Lelepa) with real PARADISEC data, and the numbers show a clear bind: full fine-tuning holds prior knowledge better but learns the new language poorly, while LoRA (and DoRA/O-LoRA) learns the target faster yet forgets earlier languages hard. That sequential Nafsan-to-Lelepa result in Table 5 is the new piece; the layer-wise cosine drift plots simply document that Lelepa forces early-encoder restructuring that the others do not.

What they do well is straightforward and useful. They vary data volume from 0.5 h to 10 h, compare full FT against encoder-decoder LoRA and the two regularized variants, and report both target CER/WER and English forgetting on the same checkpoints. The corpus stats and the component-ablation in Table 4 are clean enough that you can see the encoder is where the damage happens. Citations are ordinary and the self-cites are to their own prior low-resource work, not circular. No invented math; everything is measured against held-out sets.

Soft spots are ordinary for this kind of empirical ASR note and do not sink the claim. Standard deviations are promised for three seeds but never appear in the tables. The drift metric is only correlational; they never ablate early-encoder freezes or match update magnitude to show it causes the forgetting. The sequential pattern also reverses the usual PEFT stability advantage without a high-resource control sequence, so the “linguistic distance forces the dilemma” story is under-supported even though the raw numbers are consistent. Only one base model, no public code. None of that makes the measurements fake.

This is for anyone working on low-resource or Indigenous ASR who needs a concrete warning that the standard LoRA recipe does not transfer cleanly. It deserves a serious referee; the gaps are fixable reporting and control issues, not conceptual collapse. I would engage with it and expect it to appear after revision.

Referee Report

4 major / 6 minor

Summary. The paper presents an empirical study of adapting Whisper-Small to three low-resource Pacific Indigenous languages (Bislama, Nafsan, Lelepa) drawn from PARADISEC. It measures cross-lingual transfer under varying data budgets (0.5–10 h) with full fine-tuning versus LoRA, quantifies layer-wise representational drift via cosine distance of hidden states (min-max normalised per language), and evaluates catastrophic forgetting both on high-resource languages (English/Chinese/French) and in a sequential Nafsan→Lelepa continual-learning setting that also includes DoRA and O-LoRA. The central claim is that linguistic distance induces severe internal drift, producing a strict plasticity–stability dilemma in which LoRA is initially plastic yet forgets prior languages, while full fine-tuning is more stable but less plastic.

Significance. If the claimed drift–forgetting link holds, the work supplies a concrete stress test of speech foundation models on typologically distant, genuinely low-resource Pacific data and documents a reversal of the usual PEFT stability advantage. The curated multi-language corpus, systematic data-volume sweeps, layer-wise drift profiles (Figs. 1–2), and side-by-side comparison of Full FT / LoRA / DoRA / O-LoRA (Tables 2–5) are valuable empirical contributions for the low-resource ASR and continual-learning communities. The paper correctly flags that current adaptation recipes leave the dilemma unresolved and motivates architecture-aware or dynamic methods.

major comments (4)

§3.2 and Figs. 1–2: The cosine-distance drift (after per-language min-max normalisation on a fixed multi-language set) is presented as evidence of the structural reorganisation that causes forgetting. No experiment shows that higher early-encoder drift (Lelepa) predicts larger English or prior-language degradation after controlling for update magnitude or target CER. Without such a control, ablation (e.g., freezing early encoder layers), or matched high-resource sequential baseline, the causal chain “linguistic distance → drift → plasticity–stability dilemma” remains correlational.
Table 5 (Nafsan→Lelepa): Full FT retains Nafsan WER ≈45–47 while LoRA/DoRA/O-LoRA jump to ≈84–87, reversing both the usual PEFT stability advantage and the pattern in Table 3 (where Full FT forgets English more severely). The manuscript offers no explanation, matched same-language sequential control, or high-resource sequential control that would isolate linguistic distance from other factors (data volume imbalance, optimisation trajectory, vocabulary extension). This reversal is load-bearing for the claimed dilemma and needs either mechanistic analysis or additional controls.
§3.4 states that three random seeds are run and mean/std of CER/WER are reported, yet Tables 2–5 contain only point estimates with no standard deviations or confidence intervals. Without variance, it is impossible to judge whether the Full-FT vs LoRA differences (especially the Lelepa 2 h crossover and the sequential forgetting gaps) are reliable.
Table 3 caption claims evaluation on “English LibriSpeech” yet the table columns report Chinese CER, English WER and French WER; the experimental protocol for the non-English languages is never described. Clarify the evaluation sets and ensure the forgetting numbers are comparable across languages.

minor comments (6)

§2 opens with “four underrepresented Pacific Indigenous languages” but only three (Bislama, Nafsan, Lelepa) are listed and tabulated; correct the count.
Table 1 header uses “locales”; the body and rest of the paper use language names—align terminology.
LoRA rank, alpha and target modules are never stated; add them for reproducibility.
Figures 1–2 lack error bars or indication of which checkpoint (Full FT vs LoRA, which data budget) is plotted; the caption should specify.
Abstract and §1 claim “three distinct Pacific Indigenous languages” while §2 mentions four; keep consistent.
Minor typographical issues: “plasticity and stability dilemma” vs “plasticity–stability”; “Avg. WER” column in Table 5 appears misaligned with the two-stage numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: purely empirical measurements of CER/WER, cosine drift, and sequential forgetting on held-out data; self-citations are non-load-bearing background.

full rationale

The paper is an empirical study of Whisper-Small fine-tuning (full FT vs LoRA/DoRA/O-LoRA) on three Pacific languages. All reported quantities—test CER/WER under varying hours (Table 2), layer-wise cosine distances after min-max normalisation (Figs. 1–2, §3.2), English forgetting (Tables 3–4), and Nafsan→Lelepa sequential WERs (Table 5)—are measured against held-out splits and an external English reference. None is algebraically forced by a definition, a fitted parameter renamed as a prediction, or a self-cited uniqueness theorem. Author self-citations ([17],[18],[19],[22],[32],[33]) appear only as related prior work on keyword spotting or continual learning and do not underwrite the central measurements or the plasticity-stability claim. The causal narrative (linguistic distance → drift → dilemma) is interpretive, not circular by construction. Score 0 is therefore required under the hard rules.

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

The paper is an empirical adaptation study; it inherits the Whisper architecture, the LoRA/DoRA/O-LoRA formulations and standard ASR metrics without introducing new theoretical objects. Free parameters are the usual training hyper-parameters and the data-volume budgets chosen for the learning curves. Domain assumptions concern the linguistic distance of the three Pacific languages and the validity of cosine drift as a proxy for representational change. No new physical or mathematical entities are postulated.

free parameters (3)

peak learning rate = 1e-4
Set by hand to 1e-4 for all runs; controls the magnitude of representational change and therefore the observed drift and forgetting.
data-volume budgets (0.5/1/2/5/10 h) = 0.5, 1, 2, 5, 10 hours
Hand-chosen supervision levels used to construct the learning curves in Table 2; the qualitative conclusions about data-volume effects depend on these discrete points.
LoRA rank / alpha (implicit)
Exact rank and scaling factor are not stated; they determine the plasticity of the adapter and therefore the forgetting numbers in Tables 4–5.

assumptions (3)

domain assumption Whisper-Small’s pretrained multilingual representations are a suitable starting point for measuring cross-lingual drift on Pacific languages.
Invoked throughout §3; if the base model already lacks the necessary acoustic primitives, the measured drift is an artefact of model choice rather than of linguistic distance.
ad hoc to paper Cosine distance of hidden states (min-max normalised per language) quantifies the structural reorganisation responsible for catastrophic forgetting.
Defined in §3.2 and used to interpret Figures 1–2; no independent validation that this metric predicts forgetting magnitude is supplied.
domain assumption The three chosen languages (Bislama, Nafsan, Lelepa) are sufficiently representative of “linguistically distant Pacific Indigenous languages.”
Stated in §2 and the abstract; generalisation of the plasticity-stability dilemma rests on this sampling claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Continual Adaptation for Pacific Indigenous Speech Recognition." pith.science (2026). https://pith.science/paper/UUR6K2MD

@misc{pith2026260306310,
  author       = {Pith},
  title        = {Pith review of: Continual Adaptation for Pacific Indigenous Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UUR6K2MD}},
  note         = {Machine review of arXiv:2603.06310}
}

read the original abstract

Speech foundation models struggle with low-resource Pacific Indigenous languages because of severe data scarcity. Furthermore, full fine-tuning risks catastrophic forgetting. To address this gap, we present an empirical study adapting models to real-world Pacific datasets. We investigate the impact of data volume, adaptation strategies, and representational drift on speech foundation models for various Pacific languages. Additionally, we analyze a continual learning framework for sequential language acquisition. Empirical results across three distinct Pacific Indigenous languages demonstrate that adapting to these linguistically distant languages induces severe internal representational drift. Consequently, these models face a strict plasticity and stability dilemma. While LoRA adapts well initially, it suffers from catastrophic forgetting during sequential learning. Ultimately, this study highlights the urgent need for robust adaptation strategies tailored to underrepresented languages.

Figures

Figures reproduced from arXiv: 2603.06310 by the authors.

**Figure 2.** Relative representational drift across the twelve decoder layers. Furthermore, the table highlights a critical performance shift between the two adaptation strategies. For Bislama and Nafsan, full fine-tuning consistently achieves lower error rates than Low-Rank Adaptation across all time settings. Conversely, Lelepa presents a unique scenario under extreme data scarcity. At 0.5 and 1.0 hours, full fine-tuning stil… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 2 linked inside Pith

[1]

Introduction Despite rapid progress in automatic speech recognition (ASR) [1, 2], the vast majority of the world’s languages remain excluded from modern speech technologies [3]. This dispar- ity is particularly pronounced for Pacific languages [4, 5, 6], many of which are spoken by relatively small communities, ex- hibit limited standardized orthographies...

arXiv 2026
[2]

This dataset encompassesBislama[23],Nafsan[24], andLelepa[25]

Pacific Indigenous Speech Corpus To thoroughly evaluate our adaptation strategies under real- world conditions, we introduce a newly curated speech cor- pus comprising four underrepresented Pacific Indigenous lan- guages. This dataset encompassesBislama[23],Nafsan[24], andLelepa[25]. This selection offers diverse linguistic fea- tures and varying resource...
[3]

Methodology and Experimental Setup 3.1. Cross-lingual Transfer To analyse the effectiveness of cross-lingual transfer, we adapt a multilingual speech foundation model pretrained on high- resource languages to each Pacific language in our corpus. Specifically, we fine-tune the Whisper-Small [3] model on Bislama, Nafsan, and Lelepa with progressively increa...
[4]

Results and Analysis 4.1. Cross-Lingual Adaptation Effectiveness To quantify cross-lingual transfer under limited supervision, we evaluate the model performance across different training dura- tions and fine-tuning methods. Table 2 details the error rates for Bislama, Nafsan, and Lelepa. First, the results reveal a direct relationship between data volume ...
[5]

First, we contradict the assumption of universal adaptability

Discussion and Conclusion Synthesizing our results establishes guidelines for deploying speech models in Pacific communities. First, we contradict the assumption of universal adaptability. Adapting to low-resource languages like Lelepa and Nafsan causes severe catastrophic forgetting. This failure occurs because their vast linguistic dis- tance forces mod...
[6]

Generative AI Use Disclosure We use generative AI tools for polishing the manuscript, e.g., correcting the grammar
[7]

Automatic speech recognition for under-resourced languages: A survey,

L. Besacier, E. Barnard, A. Karpov, and T. Schultz, “Automatic speech recognition for under-resourced languages: A survey,” Speech communication, vol. 56, pp. 85–100, 2014

2014
[8]

A survey of multilingual models for automatic speech recognition,

H. Yadav and S. Sitaram, “A survey of multilingual models for automatic speech recognition,” inProceedings of the Thirteenth Language Resources and Evaluation Conference, 2022, pp. 5071– 5079

2022

Show all 39 references

[9]

Robust speech recognition via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

2023
[10]

Char- acterization of speech similarity between australian aboriginal and high-resource languages: A case study on dharawal,

T. Dang, T. M. Jeyaseelan, E. Ambikairajah, and V . Sethu, “Char- acterization of speech similarity between australian aboriginal and high-resource languages: A case study on dharawal,” in2025 Asia Pacific Signal and Information Processing Association An- nual Summit and Confe...

2025
[11]

A study of speech embedding similarities between australian aboriginal and high- resource languages,

E. Ambikairajah, J. Wu, T. Dang, and V . Sethu, “A study of speech embedding similarities between australian aboriginal and high- resource languages,”Proc. Interspeech 2025, pp. 1498–1502, 2025

2025
[12]

Development of automatic speech recognition for the documentation of cook islands m ¯aori,

R. Coto-Solano, S. A. Nicholas, S. Datta, V . Quint, P. Wills, E. N. Powell, L. Koka’ua, S. Tanveer, and I. Feldman, “Development of automatic speech recognition for the documentation of cook islands m ¯aori,” inProceedings of the Thirteenth Language Re- sources and Evaluation...

2022
[13]

Dg-sed: Domain gen- eralization for sound event detection with heterogeneous training data,

Y . Xiao, H. Yin, J. Bai, and R. K. Das, “Dg-sed: Domain gen- eralization for sound event detection with heterogeneous training data,” in2025 Asia-Pacific Signal and Information Processing As- sociation Annual Summit and Conference (APSIPA ASC), 2025

2025
[14]

Adakws: Towards robust keyword spotting with test-time adaptation,

Y . Xiao, T. Peng, Y . Zhou, and R. K. Das, “Adakws: Towards robust keyword spotting with test-time adaptation,” inINTER- SPEECH 2025, 2025

2025
[15]

Ml- superb 2.0: Benchmarking multilingual speech models across modeling constraints, languages, and datasets,

J. Shi, S.-H. Wang, W. Chen, M. Bartelds, V . Bannihatti Kumar, J. Tian, X. Chang, D. Jurafsky, K. Livescu, H.-y. Leeet al., “Ml- superb 2.0: Benchmarking multilingual speech models across modeling constraints, languages, and datasets,” inProc. Inter- speech 2024, 2024, pp. 1230–1234

2024
[16]

Omnilingual asr: Open-source multilingual speech recognition for 1600+ languages,

A. Omnilingual, G. Keren, A. Kozhevnikov, Y . Meng, C. Rop- ers, M. Setzler, S. Wang, I. Adebara, M. Auli, C. Baliogluet al., “Omnilingual asr: Open-source multilingual speech recognition for 1600+ languages,”arXiv preprint arXiv:2511.09690, 2025

2025
[17]

Unsupervised cross-lingual representation learning at scale,

A. Conneau, K. Khandelwal, N. Goyal, V . Chaudhary, G. Wenzek, F. Guzm´an, E. Grave, M. Ott, L. Zettlemoyer, and V . Stoyanov, “Unsupervised cross-lingual representation learning at scale,” in Proceedings of the 58th annual meeting of the association for computational linguist...

2020
[18]

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

A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

2020
[19]

Multilingualism in vanuatu: Four case studies,

M. Walworth, A. Dewar, T. Ennever, L. Takau, and I. Rodriguez, “Multilingualism in vanuatu: Four case studies,”International Journal of Bilingualism, vol. 25, no. 4, pp. 1120–1141, 2021

2021
[20]

Continual learning in automatic speech recognition

S. Sadhu and H. Hermansky, “Continual learning in automatic speech recognition.” inInterspeech, 2020, pp. 1246–1250

2020
[21]

Exploration of whisper fine-tuning strategies for low-resource asr,

Y . Liu, X. Yang, and D. Qu, “Exploration of whisper fine-tuning strategies for low-resource asr,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2024, no. 1, p. 29, 2024

2024
[22]

Cl-masr: A continual learning benchmark for multilin- gual asr,

L. Della Libera, P. Mousavi, S. Zaiem, C. Subakan, and M. Ra- vanelli, “Cl-masr: A continual learning benchmark for multilin- gual asr,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, 2024

2024
[23]

Adapting where it matters: Depth-aware adaptation for efficient multilingual speech recognition in low-resource languages,

Y . Xiao, E.-J. Holden, and T. Dang, “Adapting where it matters: Depth-aware adaptation for efficient multilingual speech recognition in low-resource languages,”arXiv preprint arXiv:2602.01008, 2026

2026
[24]

Rainbow keywords: Efficient incremental learning for online spoken keyword spotting,

Y . Xiao, N. Hou, and E. S. Chng, “Rainbow keywords: Efficient incremental learning for online spoken keyword spotting,” inIN- TERSPEECH 2022, 2022

2022
[25]

Dark experience for incremental keyword spotting,

T. Peng and Y . Xiao, “Dark experience for incremental keyword spotting,”arXiv preprint:2409.08153, 2024

2024 arXiv
[26]

Overcoming catastrophic forgetting in neural networks,

J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Des- jardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska- Barwinskaet al., “Overcoming catastrophic forgetting in neural networks,”Proceedings of the national academy of sciences, vol. 114, no. 13, pp. 3521–3526, 2017

2017
[27]

A two-stage lora strategy for expanding language capabilities in multilingual asr models,

C. Y . Kwok, H. Liu, J. Q. Yip, S. Li, and E. S. Chng, “A two-stage lora strategy for expanding language capabilities in multilingual asr models,”IEEE Transactions on Audio, Speech and Language Processing, 2025

2025
[28]

Continual learning for on-device environmental sound classification,

Y . Xiao, X. Liu, J. King, A. Singh, E. S. Chng, M. D. Plumbley, and W. Wang, “Continual learning for on-device environmental sound classification,” inDCASE 2022 Workshop on Detection and Classification of Acoustic Scenes and Events, 2022

2022
[29]

Bislama corpus,

N. Thieberger, “Bislama corpus,” 2023

2023
[30]

South Efate (Vanuatu)(1996-2025),

Pacific And Regional Archive for Digital Sources in Endangered Cultures, “South Efate (Vanuatu)(1996-2025),” 1995

1996
[31]

Lelepa texts and recordings (2006-2009),

S. Lacrampe, “Lelepa texts and recordings (2006-2009),” 2017

2006
[32]

Lora: Low-rank adaptation of large language models

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.”ICLR, vol. 1, no. 2, p. 3, 2022

2022
[33]

Layer-wise analysis of a self-supervised speech representation model,

A. Pasad, J.-C. Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” in2021 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 914–921

2021
[34]

Comparative layer-wise anal- ysis of self-supervised speech models,

A. Pasad, B. Shi, and K. Livescu, “Comparative layer-wise anal- ysis of self-supervised speech models,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

2023
[35]

A continual learning survey: Defying forgetting in classification tasks,

M. De Lange, R. Aljundi, M. Masana, S. Parisot, X. Jia, A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A continual learning survey: Defying forgetting in classification tasks,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 7, pp. 3366–3385, 2021

2021
[36]

Dora: Weight-decomposed low-rank adaptation,

S.-Y . Liu, C.-Y . Wang, H. Yin, P. Molchanov, Y .-C. F. Wang, K.- T. Cheng, and M.-H. Chen, “Dora: Weight-decomposed low-rank adaptation,” inForty-first International Conference on Machine Learning, 2024

2024
[37]

Orthogonal subspace learning for lan- guage model continual learning,

X. Wang, T. Chen, Q. Ge, H. Xia, R. Bao, R. Zheng, Q. Zhang, T. Gui, and X.-J. Huang, “Orthogonal subspace learning for lan- guage model continual learning,” inFindings of the Association for Computational Linguistics: EMNLP 2023, 2023, pp. 10 658– 10 671

2023
[38]

Analyt- ickws: Towards exemplar-free analytic class incremental learning for small-footprint keyword spotting,

Y . Xiao, T. Peng, R. K. Das, Y . Hu, and H. Zhuang, “Analyt- ickws: Towards exemplar-free analytic class incremental learning for small-footprint keyword spotting,” inACL 2025, 2025

2025
[39]

Aft: An exemplar-free class incremental learning method for environmental sound clas- sification,

X. Chen, X. Chen, Z. Weng, and Y . Xiao, “Aft: An exemplar-free class incremental learning method for environmental sound clas- sification,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025

2026

Pith tools

Reviewed July 15, 2026 · model on record in the stance chip above.

[1] [1]

Introduction Despite rapid progress in automatic speech recognition (ASR) [1, 2], the vast majority of the world’s languages remain excluded from modern speech technologies [3]. This dispar- ity is particularly pronounced for Pacific languages [4, 5, 6], many of which are spoken by relatively small communities, ex- hibit limited standardized orthographies...

arXiv 2026

[2] [2]

This dataset encompassesBislama[23],Nafsan[24], andLelepa[25]

Pacific Indigenous Speech Corpus To thoroughly evaluate our adaptation strategies under real- world conditions, we introduce a newly curated speech cor- pus comprising four underrepresented Pacific Indigenous lan- guages. This dataset encompassesBislama[23],Nafsan[24], andLelepa[25]. This selection offers diverse linguistic fea- tures and varying resource...

[3] [3]

Methodology and Experimental Setup 3.1. Cross-lingual Transfer To analyse the effectiveness of cross-lingual transfer, we adapt a multilingual speech foundation model pretrained on high- resource languages to each Pacific language in our corpus. Specifically, we fine-tune the Whisper-Small [3] model on Bislama, Nafsan, and Lelepa with progressively increa...

[4] [4]

Results and Analysis 4.1. Cross-Lingual Adaptation Effectiveness To quantify cross-lingual transfer under limited supervision, we evaluate the model performance across different training dura- tions and fine-tuning methods. Table 2 details the error rates for Bislama, Nafsan, and Lelepa. First, the results reveal a direct relationship between data volume ...

[5] [5]

First, we contradict the assumption of universal adaptability

Discussion and Conclusion Synthesizing our results establishes guidelines for deploying speech models in Pacific communities. First, we contradict the assumption of universal adaptability. Adapting to low-resource languages like Lelepa and Nafsan causes severe catastrophic forgetting. This failure occurs because their vast linguistic dis- tance forces mod...

[6] [6]

Generative AI Use Disclosure We use generative AI tools for polishing the manuscript, e.g., correcting the grammar

[7] [7]

Automatic speech recognition for under-resourced languages: A survey,

L. Besacier, E. Barnard, A. Karpov, and T. Schultz, “Automatic speech recognition for under-resourced languages: A survey,” Speech communication, vol. 56, pp. 85–100, 2014

2014

[8] [8]

A survey of multilingual models for automatic speech recognition,

H. Yadav and S. Sitaram, “A survey of multilingual models for automatic speech recognition,” inProceedings of the Thirteenth Language Resources and Evaluation Conference, 2022, pp. 5071– 5079

2022