Pith. sign in

REVIEW 3 major objections 5 minor 50 references

Which Languages Transfer Best to Warlpiri? A Similarity-Based Study for Low-Resource ASR

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Acoustic and typological similarity to Warlpiri picks source languages that cut ASR error rates far below monolingual and multilingual baselines.

desk verdict Useful Warlpiri ASR ranking with a real confound: source fine-tuning is not equalized for data size or Whisper pre-training exposure. read the letter →

arxiv 2607.10256 v1 pith:TSHMUHK5 submitted 2026-07-11 cs.CL eess.AS

Pravina Mylvaganam , Eliathamby Ambikairajah , Ting Dang , Vidhyasaharan Sethu , Tuende Szalay This is my paper
classification cs.CLeess.AS
keywords low-resourceASRcross-lingualtransferlanguagesimilarityspeechembeddingsAboriginallanguagesWarlpiriWhisperfine-tuningphonemeinventory
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

Warlpiri has only about 1.5 hours of usable transcribed speech, so training a speech recognizer from scratch fails. This paper shows that the right high-resource source language can be chosen systematically rather than by family, geography, or data volume alone. It ranks candidates by two complementary measures: cosine similarity of embeddings taken from pre-trained speech models, and linguistic distances computed from phoneme inventories, syntax, grammar, and overall typology. When Whisper is first adapted to the top-ranked sources and then fine-tuned on the tiny Warlpiri set, Assamese and Hindi produce word and character error rates well below those of a monolingual Warlpiri model, the original multilingual Whisper, or models adapted to dissimilar languages. Correlation analysis further shows that acoustic embedding similarity best predicts fine-tuning success, while phoneme-inventory and typological overlap better predict zero-shot transfer. The practical result is a concrete, multi-level recipe for selecting transfer sources when almost no target-language data exist.

What carries the argument

A dual similarity ranking that combines (1) average cosine similarity of utterance embeddings extracted from pre-trained speech models (ECAPA-TDNN, wav2vec 2.0, layer-wise XLSR-53) with (2) cosine/Hamming distances on linguistic feature vectors (phoneme inventory, syntax, grammar, typology), then uses the ranking to choose which high-resource language to adapt Whisper on before fine-tuning on Warlpiri.

What would settle it

Repeat the same source-language ranking and Whisper fine-tuning pipeline on another extremely low-resource language whose true best transfer sources are already known by independent phonetic or ASR evidence; if the embedding-plus-typology ranking systematically fails to recover those known best sources, the central claim is falsified.

Watch

Extended reading notes

Core claim

Acoustically and typologically similar high-resource languages, especially Assamese and Hindi, yield substantially lower word and character error rates on Warlpiri ASR after Whisper fine-tuning than monolingual training, the original multilingual Whisper, XLSR-53, or dissimilar sources; acoustic embedding similarity is the strongest predictor of fine-tuning performance, while phoneme-inventory and typological similarity better explain zero-shot transfer.

Load-bearing premise

The claim rests on the premise that cosine similarity of embeddings from models never trained on Warlpiri is a faithful proxy for the acoustic and phonetic affinity that actually drives ASR transfer success.

Editorial extensions

If this is right

  • Source-language selection for low-resource ASR can be automated by computing acoustic embedding similarity and linguistic feature distances rather than relying on language family or geography.
  • When even a few minutes of target data are available for fine-tuning, acoustic similarity should be weighted most heavily; when no target data exist (zero-shot), phoneme-inventory and typological overlap become the better guides.
  • The same dual-ranking procedure can be applied immediately to other Australian Aboriginal languages that share Warlpiri’s data-scarcity profile.
  • Embedding-layer analysis can further narrow candidate sources by revealing whether similarity is concentrated in low-level acoustic or mid-level phonetic layers.

Reading between the lines

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

  • If the embedding geometry mainly encodes speaker or channel factors rather than transferable phonetics, the ranking would still look coherent yet fail to generalize, so future work should ablate speaker-normalized embeddings.
  • The method implies that genealogically distant languages can still be optimal transfer sources whenever their acoustic and inventory profiles align, challenging family-based transfer defaults still common in speech technology.
  • A natural next test is whether the same ranking predicts gains for other speech tasks (keyword spotting, speaker diarization) on the same low-resource languages.
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

3 major / 5 minor

Summary. The paper studies cross-lingual ASR transfer to Warlpiri (~1.5 h DoReCo data) by ranking high-resource sources with two families of similarity: (i) cosine similarity of utterance embeddings from ECAPA-TDNN, wav2vec 2.0 and layer-wise XLSR-53 (Eqs. 1–2, Figs. 1–2), and (ii) linguistic distances over syntax, phoneme inventory, grammar and typology from WALS/SSWL/PHOIBLE/Grambank (Fig. 3). Candidates are first filtered by ECAPA-TDNN LID on VoxLingua107 (§2.1). Whisper-small is then adapted on each source and fine-tuned on Warlpiri; Assamese and Hindi yield the lowest WER/CER (Table 1: 32.6/12.3 and 37.6/14.3), beating monolingual (86.9/41.3), multilingual Whisper (41.0/15.1), XLSR-53 and dissimilar sources. Spearman correlations (Table 2) indicate acoustic similarity best predicts fine-tuning performance while inventory/typology better predict zero-shot transfer.

Significance. If the ranking is driven by the claimed similarities rather than confounds, the work supplies a concrete, multi-level recipe for source selection for extremely low-resource and Indigenous ASR, where genealogical or geographic heuristics are known to be unreliable. The combination of embedding-based acoustic measures with publicly documented linguistic features, the explicit zero-shot vs fine-tuning correlation split, and the focus on a typologically distant Pama–Nyungan language are useful contributions. Strengths include a clear experimental pipeline, public linguistic databases, and transparent reporting of both best and worst sources. The practical impact for documentation communities is real provided the confounds identified below are controlled.

major comments (3)
  1. §3.2 and Table 1: Intermediate source-language fine-tuning is not controlled for data quantity, domain, or Whisper pre-training exposure. The paper never states how many hours of Assamese/Hindi/Tamil/etc. were used, nor whether those languages already appear in Whisper’s original mixture (Assamese/Hindi almost certainly do). Without equalized source hours and an explicit check of pre-training coverage, the WER/CER ordering and the acoustic-similarity correlations in Table 2 can be explained by residual pre-training exposure or corpus size rather than by Eqs. 1–2 or the linguistic distances of §2.3. This is load-bearing for the central claim.
  2. §2.1–2.2 and Table 2: Candidate selection and acoustic ranking both rely on related embedding models (ECAPA-TDNN LID probabilities, then cosine similarity of ECAPA/wav2vec/XLSR embeddings). The mild circularity is acknowledged by the authors’ inclusion of English/Japanese, but the Spearman coefficients are computed on only ~8–11 languages. With such small N the reported ρ values (e.g., acoustic ρ_WER = −0.67) are unstable; confidence intervals or leave-one-out sensitivity should be supplied before claiming that acoustic similarity is “the strongest predictor.”
  3. §4.3 and Table 1: The monolingual baseline (WER 86.9 %) is trained from the same Whisper-small checkpoint on only ~1 h of Warlpiri. A stronger control would be a randomly initialized or English-only Whisper fine-tuned under identical schedules, and/or multi-source mixtures of similar vs dissimilar languages. Without these, it remains unclear how much of the gain is simply “any high-resource adaptation” versus the specific similarity ranking.
minor comments (5)
  1. §3.1: Exact train/val/test speaker partitioning and whether any speaker appears in more than one split should be stated; with only 18 speakers leakage risk is non-trivial.
  2. Fig. 3 caption and §2.3: Clarify whether cosine or Hamming is used for the final ranking that feeds Table 1; the two matrices are not identical.
  3. §4.2: Assamese and Japanese lack full Grambank/WALS coverage; the paper should note how partial feature vectors affect the typology column of Table 2.
  4. Typos / style: “Wave2vec” vs “wav2vec” inconsistency in Fig. 1 caption; “M¯aori” encoding; arXiv ID 2607.10256 appears future-dated—verify metadata.
  5. §3.2: Learning-rate schedule (10−5 after 10 % warm-up) and checkpoint selection every 200 steps are fine, but total steps / effective epochs on the 1 h set should be reported for reproducibility.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity: similarities are computed independently of the Whisper ASR evaluation; only a non-essential self-citation for LID pre-filtering exists.

  1. self citation load bearing [§2.1 (LID-based similarity) and Introduction]
    "Candidate languages are first identified using the LID-based similarity approach [26], which selects languages proximate to Warlpiri... A recent study [26] examined similarity between Warlpiri and several high-resource languages using a language identification framework."

    The pre-selection of the 11 languages that later receive full acoustic/linguistic ranking and ASR transfer experiments relies exclusively on a prior paper by overlapping authors that itself used ECAPA-TDNN embeddings on the same VoxLingua107 data. This creates a mild dependence between the candidate pool and the subsequent embedding-based acoustic scores, although the decisive Whisper WER/CER numbers and Spearman correlations remain independent measurements and the self-citation is not used to underwrite the performance claims.

full rationale

The paper's central claims (Assamese/Hindi yield lowest WER/CER after source-then-Warlpiri fine-tuning of Whisper; acoustic cosine similarity is the strongest Spearman predictor of fine-tuning performance while inventory/typology better predict zero-shot) rest on independent downstream measurements. Acoustic similarity is defined by cosine of utterance embeddings (Eqs. 1-2) extracted from ECAPA-TDNN/wav2vec/XLSR-53 never trained on Warlpiri; linguistic distances are cosine/Hamming on feature vectors from WALS/PHOIBLE/Grambank. These rankings select sources, after which Whisper is fine-tuned and evaluated on held-out Warlpiri WER/CER (Table 1); Spearman correlations are then computed post-hoc (Table 2). Nothing is fitted to Warlpiri ASR data and then re-presented as a prediction, nor is any performance number definitional from the similarity scores. The sole self-citation ([26], overlapping authors) supplies only the LID pre-filter that narrows the 107-language pool to 11 candidates; it is not invoked to justify the transfer results or correlations, which are new experiments. Confounds such as unequal source-data volume or Whisper pre-training exposure are validity issues, not circular reductions. The derivation chain is therefore empirical and self-contained.

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

The central claim rests on standard speech-embedding geometry, public typological databases, and a conventional Whisper fine-tuning protocol. No new physical entities are postulated. Free parameters are ordinary training and sampling choices; the load-bearing modeling assumptions are that embedding cosine and database feature vectors are valid proxies for transferability.

free parameters (4)
  • Whisper fine-tuning learning rate = 1e-5
    Fixed at 1e-5 after 10 % warm-up; not swept or justified beyond Hugging Face defaults; affects final WER ranking.
  • Training epochs / batch size = 10 epochs, batch 2
    10 epochs, batch size 2 chosen for the 1 h Warlpiri set; small changes could alter relative source-language ordering.
  • High-resource subset size = 2500 utterances
    2 500 utterances (~10 h) per language sampled from VoxLingua107 for embedding similarity; arbitrary computational cutoff.
  • LID top-k cutoff = top 9
    Top 9 of 107 languages retained as candidates; the cutoff is hand-chosen and shapes the later ranking pool.
assumptions (4)
  • domain assumption Cosine similarity of fixed embeddings from ECAPA-TDNN / wav2vec 2.0 / XLSR-53 is a valid proxy for acoustic/phonetic affinity that predicts ASR transfer success.
    Invoked throughout §2.2 (Eqs. 1–2) and used to rank sources; never independently validated outside the same embedding models.
  • domain assumption Feature vectors drawn from WALS, SSWL, Ethnologue, PHOIBLE and Grambank, with missing values simply dropped, yield reliable syntactic, inventory, grammatical and typological distances.
    §2.3; missing-feature policy is stated but completeness of the databases for Warlpiri and the 11 sources is assumed.
  • domain assumption Full-model fine-tuning of Whisper-small for 10 epochs under the stated schedule is a fair and sufficient protocol for comparing source languages.
    §3.2; alternative adapters, LoRA or different schedules are not tested.
  • domain assumption The 1 h / 15 min / 15 min Warlpiri split and the removal of noisy segments produce an unbiased evaluation of transfer quality.
    §3.1; extremely small test set makes ranking sensitive to split choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Which Languages Transfer Best to Warlpiri? A Similarity-Based Study for Low-Resource ASR." pith.science (2026). https://pith.science/paper/TSHMUHK5

@misc{pith2026260710256,
  author       = {Pith},
  title        = {Pith review of: Which Languages Transfer Best to Warlpiri? A Similarity-Based Study for Low-Resource ASR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSHMUHK5}},
  note         = {Machine review of arXiv:2607.10256}
}
read the original abstract

This paper investigates how language similarity can improve cross-lingual transfer for automatic speech recognition (ASR) in extremely low-resource settings. Warlpiri, an Australian Aboriginal language, has very limited transcribed speech data, making transfer learning essential. We propose a framework combining acoustic similarity from pre-trained speech models with linguistic similarity based on typology, phoneme inventories, grammatical, and syntactic features to rank high-resource source languages and evaluate their effectiveness for ASR transfer to Warlpiri. Experiments with Whisper show that acoustically and typologically similar languages outperform monolingual and multilingual baselines. Assamese and Hindi achieve substantial reductions in word and character error rates. Correlation analysis further indicates that acoustic similarity is the strongest predictor of fine-tuning performance, while phoneme inventory and typological similarity better explain zero-shot transfer.

Figures

Figures reproduced from arXiv: 2607.10256 by the authors.

Figure 1
Figure 1. Embedding-based cosine similarity between Warlpiri and selected high-resource languages using ECAPA-TDNN and Wave2vec 2.0. were selected to span a broad range of similarity levels and language families for controlled comparison. However, due to missing linguistic feature data for Assamese and Japanese, only syntactic and phoneme inventory distances could be analyzed for these two languages. 4.2. Language Similarity … view at source ↗
Figure 2
Figure 2. Cosine similarity between Warlpiri and selected high￾resource languages across layers of XLSR-53 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The language similarity matrix measures cosine (left) and hamming (right) similarity between Warlpiri and selected high-resource languages. However, some discrepancies exist; for instance, Maori shows ¯ higher syntactic similarity despite lower acoustic proximity, highlighting the importance of considering multiple similarity dimensions. By integrating acoustic and linguistic-feature anal￾yses, we gain a more compre… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 7 linked inside Pith

  1. [1]

    In recent years, ASR has advanced markedly due to large-scale annotated datasets and powerful deep learning models [1]

    Introduction Automatic Speech Recognition (ASR) has become a transfor- mative technology that enables natural human–computer inter- action in applications such as virtual assistants, language learn- ing, and accessibility tools. In recent years, ASR has advanced markedly due to large-scale annotated datasets and powerful deep learning models [1]. However,...

    arXiv 2026

  2. [2]

    Similarity Analysis The detailed similarity analysis between Warlpiri and the se- lected high-resource languages is presented in this section. Can- didate languages are first identified using the LID-based sim- ilarity approach [26], which selects languages proximate to Warlpiri, providing a principled basis for in-depth similarity analysis while ensuring...

  3. [3]

    Datasets Warlpiri speech recordings and their corresponding transcrip- tions were obtained from the DoReCo dataset [38]

    Experiments 3.1. Datasets Warlpiri speech recordings and their corresponding transcrip- tions were obtained from the DoReCo dataset [38]. The record- ings were first pre-processed by removing low-quality segments and utterances containing excessive noise or irrelevant speech. The remaining audio signals were then downsampled to 16 kHz to match the input r...

  4. [4]

    Results and Discussion 4.1. Candidate Language Selection Based on the LID-based pre-selection in Section 2.1, the top nine of 107 languages were selected as acoustically similar can- didates: Assamese, Hindi, Tamil, Telugu, Malayalam, Finnish, M¯aori, Javanese, and Swahili, as they show higher similarity than the others. Notably, these languages are neith...

  5. [5]

    This enables examination of how Warlpiri’s similarity to high-resource languages evolves across model layers

    indicates that lower layers capture acoustic features, mid- dle layers encode phonetic information, and higher layers rep- resent task-specific characteristics. This enables examination of how Warlpiri’s similarity to high-resource languages evolves across model layers. Among the 11 languages, Assamese and Hindi show consistently high similarity with Warl...

  6. [6]

    We presented a systematic framework that combines embedding-based acoustic similarity with linguistic- feature-based similarity to guide source language selection

    Conclusion This study examined the role of language similarity in cross- lingual ASR for Warlpiri, an extremely low-resource Australian Aboriginal language. We presented a systematic framework that combines embedding-based acoustic similarity with linguistic- feature-based similarity to guide source language selection. Ex- perimental results show that sou...

  7. [7]

    Ethics ap- proval was granted by the Human Research Ethics Advisory Panel Executive (HREAP) at UNSW Sydney, Australia, under Reference Number iRECS6867

    Acknowledgment The authors would like to thank the School of Electrical Engi- neering and Telecommunications at UNSW Sydney, Australia, for providing funding for this research initiative. Ethics ap- proval was granted by the Human Research Ethics Advisory Panel Executive (HREAP) at UNSW Sydney, Australia, under Reference Number iRECS6867

  8. [8]

    Af- ter using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the pub- lication

    Generative AI use disclosure During the preparation of this work, the authors used an AI tool in order to refine the academic language, improve the structural flow of the manuscript, and optimise technical terminology. Af- ter using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the pub- lication

Show all 50 references
  1. [9]

    End-to-end speech recognition: A survey,

    R. Prabhavalkar, T. Hori, T. N. Sainath, R. Schl¨uter, and S. Watan- abe, “End-to-end speech recognition: A survey,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 325–351, 2023

    2023

  2. [10]

    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

  3. [11]

    Language and culture: Socialisation in a warlpiri com- munity,

    E. Bavin, “Language and culture: Socialisation in a warlpiri com- munity,”Language and culture in Aboriginal Australia, pp. 85– 96, 1993

    1993

  4. [12]

    Cross-lingual transfer learn- ing during supervised training in low resource scenarios

    A. Das and M. Hasegawa-Johnson, “Cross-lingual transfer learn- ing during supervised training in low resource scenarios.” inIN- TERSPEECH, 2015, pp. 3531–3535

    2015

  5. [13]

    Cross-lingual transfer learning for low- resource speech translation,

    S. Khurana, N. Dawalatabad, A. Laurent, L. Vicente, P. Gimeno, V . Mingote, and J. Glass, “Cross-lingual transfer learning for low- resource speech translation,” in2024 IEEE International Con- ference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 2024,...

    2024

  6. [14]

    A balanced data approach for evaluating cross-lingual transfer: Mapping the lin- guistic blood bank,

    D. Malkin, T. Limisiewicz, and G. Stanovsky, “A balanced data approach for evaluating cross-lingual transfer: Mapping the lin- guistic blood bank,” inProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langua...

    2022

  7. [15]

    Low resource asr: The surprising effectiveness of high resource transliteration

    S. Khare, A. R. Mittal, A. Diwan, S. Sarawagi, P. Jyothi, and S. Bharadwaj, “Low resource asr: The surprising effectiveness of high resource transliteration.” inInterspeech, 2021, pp. 1529– 1533

    2021

  8. [16]

    Towards scalable efficient on-device asr with transfer learning,

    L. Pandey, K. Li, J. Guo, D. Paul, A. Guo, J. Mahadeokar, and X. Zhang, “Towards scalable efficient on-device asr with transfer learning,”arXiv preprint arXiv:2407.16664, 2024

    2024 arXiv

  9. [17]

    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

  10. [18]

    Exploiting adapters for cross-lingual low-resource speech recognition,

    W. Hou, H. Zhu, Y . Wang, J. Wang, T. Qin, R. Xu, and T. Shinozaki, “Exploiting adapters for cross-lingual low-resource speech recognition,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 317–329, 2021

    2021

  11. [19]

    Mutual intelligibility between closely related languages in europe,

    C. Gooskens, V . J. Van Heuven, J. Golubovi ´c, A. Sch ¨uppert, F. Swarte, and S. V oigt, “Mutual intelligibility between closely related languages in europe,”International Journal of Multilin- gualism, vol. 15, no. 2, pp. 169–193, 2018

    2018

  12. [20]

    Comparing germanic, romance and slavic: Relationships among linguistic distances,

    W. Heeringa, C. Gooskens, and V . J. Van Heuven, “Comparing germanic, romance and slavic: Relationships among linguistic distances,”Lingua, vol. 287, p. 103512, 2023

    2023

  13. [21]

    Investigating the impact of cross- lingual acoustic-phonetic similarities on multilingual speech recognition,

    M. Umar Farooq and T. Hain, “Investigating the impact of cross- lingual acoustic-phonetic similarities on multilingual speech recognition,”arXiv e-prints, pp. arXiv–2207, 2022

    2022

  14. [22]

    Cross- lingual transfer for speech processing using acoustic language similarity,

    P. Wu, J. Shi, Y . Zhong, S. Watanabe, and A. W. Black, “Cross- lingual transfer for speech processing using acoustic language similarity,” in2021 IEEE Automatic Speech Recognition and Un- derstanding Workshop (ASRU). IEEE, 2021, pp. 1050–1057

    2021

  15. [23]

    Improving cross-lingual phonetic representation of low-resource languages through language sim- ilarity analysis,

    M. Kim, K. Jang, and H. Kim, “Improving cross-lingual phonetic representation of low-resource languages through language sim- ilarity analysis,” inICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  16. [24]

    Sonorant spec- tra and coarticulation distinguish speakers with different dialects,

    C. Themistocleous, V . Fyndanis, and K. Tsapkini, “Sonorant spec- tra and coarticulation distinguish speakers with different dialects,” Speech Communication, vol. 142, pp. 1–14, 2022

    2022

  17. [25]

    On the relation between the similarity of the acoustic distribution patterns of vowels and the language closeness,

    H. Harnud and Z. Xuewen, “On the relation between the similarity of the acoustic distribution patterns of vowels and the language closeness,”International Journal of Anthropology and Ethnology, vol. 5, no. 1, p. 14, 2021

    2021

  18. [26]

    Choosing transfer lan- guages for cross-lingual learning,

    Y .-H. Lin, C.-Y . Chen, J. Lee, Z. Li, Y . Zhang, M. Xia, S. Ri- jhwani, J. He, Z. Zhang, X. Maet al., “Choosing transfer lan- guages for cross-lingual learning,” inProceedings of the 57th An- nual Meeting of the Association for Computational Linguistics, 2019, pp. 3125–3135

    2019

  19. [27]

    E. L. Bavin and E. L. Bavin,Warlpiri in the 80s: An overview of research into language variation and child language, 1987

    1987

  20. [28]

    Explaining typological differences between lan- guages: de facto topicalisation in english and warlpiri,

    T. Shopen, “Explaining typological differences between lan- guages: de facto topicalisation in english and warlpiri,” 2001

    2001

  21. [29]

    Contrasting warlpiri and english language fea- tures,

    D. M. Bryant, “Contrasting warlpiri and english language fea- tures,”International Journal of Culture and History, vol. 8, no. 2, pp. 1–15, 2021

    2021

  22. [30]

    Warlpiri and the grammar of non-configurational lan- guages,

    K. Hale, “Warlpiri and the grammar of non-configurational lan- guages,”Natural language & linguistic theory, vol. 1, no. 1, pp. 5–47, 1983

    1983

  23. [31]

    V owel harmony, directionality and mor- pheme structure constraints in warlpiri,

    M. Harvey and B. Baker, “V owel harmony, directionality and mor- pheme structure constraints in warlpiri,”Lingua, vol. 115, no. 10, pp. 1457–1474, 2005

    2005

  24. [32]

    Distinguishing prosodic word and phonological word in warlpiri: Prosodic constituency in morpho- logically complex words,

    C. Pentland and M. Laughren, “Distinguishing prosodic word and phonological word in warlpiri: Prosodic constituency in morpho- logically complex words,” 2005

    2005

  25. [33]

    An acoustic and articulatory anal- ysis of focus and the word/morpheme boundary distinction in warlpiri,

    A. Butcher and J. Harrington, “An acoustic and articulatory anal- ysis of focus and the word/morpheme boundary distinction in warlpiri,” inProceedings of the 6th international seminar on speech production, 2003, pp. 19–24

    2003

  26. [34]

    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

  27. [35]

    Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    2005 arXiv

  28. [36]

    V oxlingua107: a dataset for spoken lan- guage recognition,

    J. Valk and T. Alum ¨ae, “V oxlingua107: a dataset for spoken lan- guage recognition,” in2021 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2021, pp. 652–658

    2021

  29. [37]

    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

  30. [38]

    Unsupervised cross-lingual representation learning for speech recognition,

    A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, “Unsupervised cross-lingual representation learning for speech recognition,”arXiv preprint arXiv:2006.13979, 2020

    2006 arXiv

  31. [39]

    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

  32. [40]

    A study of conceptual lan- guage similarity: comparison and evaluation,

    H. Ye, Y . Liu, and H. Sch ¨utze, “A study of conceptual lan- guage similarity: comparison and evaluation,”arXiv preprint arXiv:2305.13401, 2023

    2023 arXiv

  33. [41]

    M. S. Dryer and M. Haspelmath, Eds.,WALS Online (v2020.4). Zenodo, 2013. [Online]. Available: https://doi.org/10.5281/ zenodo.13950591

    2013

  34. [42]

    C. C and K. R, Eds.,SSWL Database of Syntactic Parameters, 2009

    2009

  35. [43]

    D. M. Eberhard, G. F. Simons, and C. D. Fennig, Eds., Ethnologue: Languages of the World, twenty-eighth ed. Dallas, Texas: SIL International, 2025. [Online]. Available: https://www.ethnologue.com/

    2025

  36. [44]

    Moran and D

    S. Moran and D. McCloy, Eds.,PHOIBLE 2.0, 2019. [Online]. Available: https://phoible.org

    2019

  37. [45]

    Grambank reveals the impor- tance of genealogical constraints on linguistic diversity and high- lights the impact of language loss,

    H. Skirg ˚ard, H. J. Haynie, D. E. Blasi, H. Hammarstr ¨om, J. Collins, J. J. Latarche, J. Lesage, T. Weber, A. Witzlack- Makarevich, S. Passmoreet al., “Grambank reveals the impor- tance of genealogical constraints on linguistic diversity and high- lights the impact of langua...

    2023

  38. [46]

    Warlpiri doreco dataset,

    C. O’Shannessy, “Warlpiri doreco dataset,” inLanguage Docu- mentation Reference Corpus (DoReCo) 2.0, F. Seifart, L. Paschen, and M. Stave, Eds. Lyon: Laboratoire Dynamique Du Langage (UMR5596, CNRS & Universit ´e Lyon 2), 2024. [Online]. Available: https://doreco.huma-num.fr/l...

    2024

  39. [47]

    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

  40. [48]

    Whispering in amharic: Fine-tuning whisper for low-resource language,

    D. K. Gete, B. Y . Ahmed, T. D. Belay, Y . A. Ejigu, S. H. Imam, A. B. Tessema, M. O. Adem, T. A. Belay, R. Geislinger, U. A. Musaet al., “Whispering in amharic: Fine-tuning whisper for low-resource language,”arXiv preprint arXiv:2503.18485, 2025

    2025 arXiv

  41. [49]

    Available: https://huggingface.co/blog/ fine-tune-whisper

    [Online]. Available: https://huggingface.co/blog/ fine-tune-whisper

  42. [50]

    Xls-r deep learning model for multilingual asr on low-resource languages: Indone- sian, javanese, and sundanese,

    P. Arisaputra, A. T. Handoyo, and A. Zahra, “Xls-r deep learning model for multilingual asr on low-resource languages: Indone- sian, javanese, and sundanese,”arXiv preprint arXiv:2401.06832, 2024

    2024 arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.