{"id":"688ee47e-ec5c-4b19-890f-499c39729deb","arxiv_id":"2607.06289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"Continual pre-training on Sinhala followed by Dhivehi fine-tuning achieves 12.89% WER, but KenLM decoding contributes 27+ absolute WER points, dwarfing the 0.61-point transfer gain.","lead":"This paper tests whether speech recognition for Dhivehi (Maldives) improves by transferring knowledge from Sinhala, a related language. The result: a modest WER improvement, but the external language model decoder contributes far more than any transfer strategy.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The headline CPT result (12.89% WER) uses XLS-R while all other experiments use Wav2Vec2-BERT, confounding the transfer-strategy comparison. Additionally, the Turkish control does not cover the CPT paradigm, leaving the best result without a relatedness control.","rationale":"The reader correctly identified the architectural confound (XLS-R vs Wav2Vec2-BERT) as the most load-bearing concern. This is the right target: the paper's headline claim ranks CPT as the best transfer strategy, but CPT is the only experiment using a different base model, making the comparison non-controlled. I add two observations that strengthen the reader's concern: (1) the Turkish control does not cover the CPT paradigm, so the relatedness claim for the best result rests on indirect evidence from a different setup; (2) the effect sizes (0.24–0.61 WER) are within plausible single-run variance given no repeated runs.\n\nThe reader's verdict of CONDITIONAL with MODERATE confidence is appropriate. The paper has genuine merits — the controlled experimental design, the language ID token ablation, and the KenLM analysis are useful contributions for an understudied language pair. But the central ranking claim ('CPT is the best transfer strategy') is not securely established due to the confound. If the architectural inconsistency were resolved and variance quantified, the paper could move to ACCEPT. The reader's recommendation is sound; I see no reason to adjust the verdict.","tokens_in":10333,"tokens_out":1925,"duration_ms":100955,"concrete_test":"Run two additional experiments: (1) CPT on Wav2Vec2-BERT using the same Sinhala data and fine-tuning on Dhivehi, to isolate the transfer strategy from the architecture; (2) all experiments (including CPT) on XLS-R as the base, to provide a within-architecture comparison. If CPT no longer outperforms the Dhivehi-only baseline under either matched-architecture condition, the central claim weakens. Additionally, run at least 3 seeds for the top 3 configurations (baseline, multilingual no-LID, CPT) and report mean ± std; if the 0.61 WER gap falls within one standard deviation, the ranking is not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that continual pre-training (CPT) on Sinhala is the best transfer strategy for Dhivehi ASR. However, §V.E acknowledges that CPT used XLS-R while all other experiments used Wav2Vec2-BERT. This means the 0.61 WER improvement of CPT over the Dhivehi-only baseline cannot be attributed to the transfer strategy alone — XLS-R (pretrained on 128 languages, ~436K hours) may simply be a stronger base model than Wav2Vec2-BERT. The paper provides no experiment isolating architecture from transfer strategy.\n\nA second, less obvious problem compounds this: the Turkish control experiment (§III.D.6) is conducted only under the multilingual fine-tuning paradigm, not under CPT. So the claim that 'linguistic relatedness' drives the CPT improvement is supported only indirectly, via a different transfer paradigm with a different architecture. The headline result has neither an architectural control nor a relatedness control.\n\nFinally, the transfer gains are small: 0.24–0.61 absolute WER points. The paper acknowledges (§V.E) no repeated runs, no cross-validation, and no multiple data splits. In low-resource ASR, single-run variance of this magnitude is plausible. Without variance estimates, it is unclear whether any transfer strategy meaningfully outperforms the baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This paper investigates cross-lingual transfer learning from Sinhala to Dhivehi for low-resource ASR, conducting 17 experiments across five transfer paradigms (Dhivehi-only baseline, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a Turkish control). The strongest system—continual pre-training (CPT) on Sinhala followed by Dhivehi fine-tuning with KenLM decoding—achieves 12.89% WER and 2.70% CER, improving over the Dhivehi-only baseline (13.50% WER, 3.02% CER). The Turkish control experiment provides evidence that gains stem from linguistic relatedness rather than generic data augmentation. The study also finds that KenLM decoding is the single largest contributor to performance, that language ID tokens are harmful in this bilingual setting, and that multilingual fine-tuning without LID tokens is a competitive alternative to CPT.","tokens_in":11224,"tokens_out":1376,"duration_ms":199564,"significance":"The paper addresses a genuinely under-studied language pair (Sinhala–Dhivehi) and provides a controlled experimental comparison across multiple transfer paradigms, which is valuable for the low-resource ASR community. The inclusion of a Turkish unrelated-language control is a methodological strength, as is the finding that language ID tokens can be harmful in bilingual low-resource settings—a practically useful and falsifiable result. The reproducible code repository and the reporting of both WER and CER across all configurations are commendable. The KenLM effect analysis (Table IV) is a useful contribution showing that decoding configuration is a first-order effect. However, the central claim that CPT is the best transfer strategy is confounded by an architectural inconsistency (XLS-R for CPT vs. Wav2Vec2-BERT for all other experiments), and the transfer gains (0.24–0.61 absolute WER points) are reported without variance estimates, limiting the strength of the comparative conclusions.","major_comments":[{"comment":"§III.D.5 and §V.E: The CPT experiment uses XLS-R while all other experiments use Wav2Vec2-BERT. The paper acknowledges this as an 'architectural inconsistency' but does not address its implications for the central claim. The 0.61 WER improvement of CPT over the Dhivehi-only baseline cannot be cleanly attributed to the transfer strategy because XLS-R (pretrained on 128 languages, ~436K hours) may be a stronger base model than Wav2Vec2-BERT regardless of transfer. This is load-bearing for the claim that CPT is the best transfer strategy. At minimum, the paper should explicitly discuss this confound in the results interpretation (§V.A) rather than only in the limitations, and ideally should include a Dhivehi-only XLS-R baseline to isolate the architecture effect from the transfer effect.","section":null},{"comment":"§III.D.6 and §IV.E: The Turkish control experiment is conducted only under the multilingual fine-tuning paradigm, not under CPT. The paper's claim that 'linguistic relatedness' drives the CPT improvement is therefore supported only indirectly, via a different transfer paradigm with a different architecture. The headline CPT result has neither an architectural control (same base model without Sinhala transfer) nor a relatedness control (CPT with Turkish). This should be explicitly acknowledged as a limitation of the relatedness claim, or a CPT Turkish control should be added.","section":null},{"comment":"§V.E and Table III: All results are from single runs with no repeated runs, no cross-validation, and no multiple data splits. The transfer gains are small (0.24–0.61 absolute WER points). In low-resource ASR, single-run variance of this magnitude is plausible. Without any variance estimates or confidence intervals, it is unclear whether any transfer strategy meaningfully outperforms the baseline. The paper should either provide variance estimates from repeated runs or explicitly temper the comparative claims to reflect this uncertainty.","section":null}],"minor_comments":[{"comment":"Abstract: The second half is near-duplicated, repeating the list of five paradigms and the best-system description almost verbatim. This should be condensed.","section":null},{"comment":"§III.B: The text says 'Wav2Vec was selected as the primary ASR model' but the experiments use Wav2Vec2-BERT (and XLS-R for CPT). The model naming should be stated consistently and precisely.","section":null},{"comment":"§III.C: The KenLM hyperparameters (α=0.5, β=1.0, beam=64) are described as 'empirically tuned' but the search space and tuning procedure are not reported. A brief description of the tuning range and selection criterion would improve reproducibility.","section":null},{"comment":"Table III: The 'Cont. pretrain Si→Dv' row should note the base model (XLS-R) explicitly in the table or its caption, so the architectural difference is visible to readers without consulting §V.E.","section":null},{"comment":"§II.A: The text mentions 'XLS-R [8] extended this framework across 53 languages, while XLS-R [7] scaled to 128 languages.' The first appears to refer to XLSR-53 (Conneau et al., 2021) and the second to XLS-R (Babu et al., 2022). The naming should distinguish XLSR-53 from XLS-R to avoid confusion.","section":null},{"comment":"§IV.G: The comparison with published benchmarks (Table VIII) notes that Ahmed [12] used Common Voice 13.0 while this work uses Common Voice 22.0. This dataset-version difference should be stated more prominently, as it affects the comparability of the 13.50% vs. 14.26% WER numbers.","section":null},{"comment":"Figure 1 is referenced but not visible in the manuscript text provided. Ensure the figure clearly shows the genealogical relationship as described.","section":null}],"recommendation":"major_revision","confidential_remarks":"The architectural confound (XLS-R for CPT, Wav2Vec2-BERT for everything else) is the most serious issue. The authors acknowledge it in §V.E but do not seem to recognize that it undermines the central comparative claim. A Dhivehi-only XLS-R baseline would substantially strengthen the paper and should be feasible with their existing computational resources. If that baseline shows XLS-R alone does not outperform Wav2Vec2-BERT, the CPT claim becomes much more convincing. The paper is otherwise a solid contribution to an under-studied area, and the Turkish control design (even if limited to one paradigm) shows good experimental instinct."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises three substantive concerns, all of which are valid in whole or in part. We address each below and describe the revisions we will make.","responses":[{"response":"The referee is correct that the architectural inconsistency between XLS-R (used for CPT) and Wav2Vec2-BERT (used for all other experiments) confounds the attribution of the CPT improvement to the transfer strategy rather than to the base model. We acknowledge this without reservation. In the current manuscript, this issue is mentioned only in §V.E (Limitations), which is insufficient given that the headline result depends on it. We will make two changes. First, we will add an explicit discussion of this confound in §V.A (Transfer Strategy Matters More Than Relatedness Alone), making clear that the 0.61 absolute WER improvement of CPT over the Dhivehi-only baseline cannot be cleanly separated from the potentially stronger base architecture. We will temper the claim from 'CPT is the best transfer strategy' to 'CPT yields the best observed performance in this study, though the architectural difference between XLS-R and Wav2Vec2-BERT prevents a clean isolation of the transfer effect from the architecture effect.' Second, regarding the Dhivehi-only XLS-R baseline: we agree this is the ideal control and will attempt to run it. However, we must be honest that computational constraints may prevent us from completing this additional experiment within the revision timeframe. If we cannot add it, we will state explicitly that this control is missing and that the CPT result should be interpreted with the architectural confound in mind. We will not claim a clean attribution that the data do not support.","revision_made":"partial","referee_comment":"CPT uses XLS-R while all other experiments use Wav2Vec2-BERT, confounding the claim that CPT is the best transfer strategy. The paper should discuss this in results interpretation, not just limitations, and ideally add a Dhivehi-only XLS-R baseline."},{"response":"This is a fair and accurate observation. The Turkish control experiment (§III.D.6, §IV.E) is conducted under the multilingual fine-tuning paradigm using Wav2Vec2-BERT, not under CPT using XLS-R. Therefore, the claim that linguistic relatedness drives the CPT improvement is supported only indirectly—via a different transfer paradigm and a different architecture. The manuscript's conclusion states that 'the Turkish control experiment confirms that observed improvements stem from linguistic relatedness,' which overstates what the evidence supports specifically for the CPT result. We will revise this in two ways. First, we will narrow the relatedness claim in the Conclusion and in §V.D to apply specifically to the multilingual fine-tuning paradigm, where the Turkish control is directly applicable. Second, we will add an explicit statement that the CPT result lacks both an architectural control (Dhivehi-only XLS-R) and a relatedness control (CPT with Turkish), and that the extension of the relatedness finding to CPT is an inference, not a directly tested claim. As with the XLS-R baseline, we will attempt to run a CPT Turkish control if computational resources permit, but we commit to the textual correction regardless of whether the additional experiment is completed.","revision_made":"partial","referee_comment":"The Turkish control is conducted only under multilingual fine-tuning, not CPT. The relatedness claim for CPT is therefore only indirectly supported. This should be acknowledged or a CPT Turkish control should be added."},{"response":"The referee is correct. The transfer gains are small—0.24 absolute WER for multilingual fine-tuning without LID tokens and 0.61 absolute WER for CPT—and these are from single runs with no repeated runs, cross-validation, or multiple data splits. In low-resource ASR with ~37 hours of training data, run-to-run variance of this magnitude is entirely plausible, and we cannot rule out that some or all of the observed differences are within noise. We will address this in two ways. First, we will add explicit hedging language throughout the results and discussion sections, stating that the observed differences are small and that without variance estimates we cannot confirm statistical significance. We will reframe the comparative claims as observed trends rather than established differences. Second, we will attempt to run at least 3–5 repeated runs for the key configurations (Dhivehi-only baseline, multilingual no-LID, and CPT) to provide standard deviations or confidence intervals. If we are able to complete these, we will report them in a revised Table III. If computational constraints prevent sufficient repeated runs, we will state this limitation explicitly and further temper the claims. We agree with the referee that the current framing is too strong for single-run results with sub-1-point differences.","revision_made":"partial","referee_comment":"All results are from single runs with no variance estimates. Transfer gains are small (0.24–0.61 absolute WER points). Without variance estimates, it is unclear whether any transfer strategy meaningfully outperforms the baseline."}],"tokens_in":10211,"tokens_out":1229,"duration_ms":73340,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this is a solid empirical study of Sinhala-to-Dhivehi cross-lingual transfer, a language pair nobody has studied before. The experimental design is thoughtful — five transfer paradigms, 17 runs, a Turkish unrelated-language control, and a clean ablation on language ID tokens. The code is public on GitHub. The central confound is real but fixable: the best result (CPT, 12.89% WER) uses XLS-R while every other experiment uses Wav2Vec2-BERT, so you can't cleanly attribute the gain to the transfer strategy versus the stronger base model. The authors acknowledge this honestly in §V.E but don't resolve it. The Turkish control also only covers the multilingual paradigm, not CPT, so the relatedness argument for the headline result is indirect. The transfer gains are small (0.24–0.61 absolute WER), and with no repeated runs or cross-validation, it's hard to know if they're real. That said, the paper earns credit on several fronts. The language ID token ablation is a genuinely useful finding — showing that LID tokens hurt in this bilingual character-level setting is counterintuitive and well-demonstrated. The KenLM effect (24–29 absolute WER points across all configs) is correctly identified as the dominant factor, and the paper is refreshingly upfront about this rather than overselling the transfer gains. The Turkish control does its job for the multilingual paradigm: Sinhala helps, Turkish hurts slightly, baseline is in between. The Dhivehi-only baseline (13.50% WER) already beats the prior published benchmark (14.26%), which makes the transfer gains meaningful rather than improvements over a weak baseline. The Sinhala baseline (7.00% WER) is a big jump over the prior 17.19%, though the authors note the comparison isn't clean due to data version differences. The stress-test concern about the architectural confound lands fully — it's the paper's main weakness and the authors know it. The concern about single-run variance is also valid but somewhat less alarming given that the effects are consistent in direction across related experiments. This paper is for low-resource ASR practitioners and researchers working on Indo-Aryan languages. It deserves a serious referee who can push the authors to either run CPT on Wav2Vec2-BERT or run all experiments on XLS-R, and to report at least 3 seeds for the key comparisons. If they do that, this becomes a clean contribution. Recommend peer review with a revision requirement on the architectural confound and variance reporting.","headline":"Sinhala-to-Dhivehi transfer study: useful empirical work on an unstudied language pair, but the headline CPT result is architecturally confounded.","tokens_in":11135,"tokens_out":614,"would_cite":false,"duration_ms":60796,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Related-Language Transfer Beats Baseline for Dhivehi ASR","keywords":["automatic speech recognition","cross-lingual transfer learning","low-resource languages","Dhivehi","Sinhala","continual pre-training","multilingual fine-tuning","language model decoding"],"falsifier":"If replacing Sinhala with Turkish in the continual pre-training setup still produced WER improvements over the Dhivehi-only baseline comparable to the Sinhala CPT result, the claim that gains stem from linguistic relatedness would be undermined.","tokens_in":10421,"feed_emoji":"🗣️","tokens_out":1294,"duration_ms":116279,"temperature":0.7,"pith_summary":"This paper tests whether speech recognition for Dhivehi, a language spoken by fewer than half a million people with only 37 hours of validated training data, can be improved by leveraging Sinhala, a related Indo-Aryan language with roughly six times more available data. The authors run seventeen experiments across five transfer strategies—Dhivehi-only baselines, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a Turkish control. Their best system, continual pre-training on Sinhala followed by Dhivehi fine-tuning with an external language model, achieves 12.89% word error rate, improving on the Dhivehi-only baseline of 13.50%. A Turkish control experiment, using an unrelated language matched in data volume, performs worse than the baseline, which the authors interpret as evidence that the gains stem from genuine linguistic relatedness rather than generic data augmentation. However, the paper also finds that the external n-gram language model contributes by far the largest improvement—roughly 28 absolute WER points across all systems—dwarfing every transfer strategy. Additionally, language identification tokens, often beneficial in large-scale multilingual ASR, hurt performance in this bilingual setting, and sequential fine-tuning failed to improve over the baseline at all.","feed_headline":"Sinhala Transfer Cuts Dhivehi Speech Errors to 12.89% WER","feed_subtitle":"Cross-lingual transfer from a related Indo-Aryan language improves low-resource Dhivehi ASR, but external language model decoding matters 40","key_machinery":"The paper compares five transfer learning paradigms applied to self-supervised speech models (Wav2Vec2-BERT and XLS-R): (1) Dhivehi-only fine-tuning as baseline, (2) sequential fine-tuning from Sinhala to Dhivehi, (3) multilingual joint fine-tuning with and without language ID tokens, (4) continual pre-training on Sinhala audio followed by Dhivehi fine-tuning, and (5) a Turkish control replacing Sinhala with an unrelated language at matched data volume. An external 5-gram KenLM language model is integrated via shallow fusion with pyctcdecode beam search decoding (beam width 64, alpha 0.5, beta 1.0).","core_discovery":"The central finding is that cross-lingual transfer from a related language to a low-resource one works, but only through specific training strategies—continual pre-training and multilingual fine-tuning without language ID tokens—while the choice of decoding configuration (an external n-gram language model) has a far larger effect on final accuracy than any transfer approach. The Turkish control experiment, which degrades performance below the baseline, is used to argue that the modest gains from Sinhala reflect genuine phonological and acoustic overlap between the two Insular Indo-Aryan languages rather than a generic benefit from adding more training data.","pith_inferences":["The architectural inconsistency between continual pre-training (XLS-R) and all other experiments (Wav2Vec2-BERT) means the best result cannot be cleanly attributed to the transfer strategy alone; a controlled comparison using the same base architecture for CPT vs. other strategies would be needed to isolate the effect.","The near-clustering of all systems at 40-44% WER without language model decoding suggests that the acoustic models alone may not be learning sufficiently discriminative representations in this data regime, and the KenLM is compensating for acoustic model weakness rather than complementing strong representations.","If the finding about language ID tokens generalizes, it could imply that explicit language conditioning is only beneficial when the model must distinguish among many languages; in a two-language setting, the added tokens may partition the decoder's capacity unnecessarily.","The 0.51 WER gap between the best transfer system (12.89%) and the Turkish control (13.77%) is small relative to the 28-point KenLM effect, raising the question of whether the relatedness signal would survive statistical significance testing across multiple random seeds or data splits."],"forward_implications":["For low-resource ASR, external language model decoding may matter more than the choice of acoustic model or transfer strategy, suggesting that n-gram LM integration should be a first-class experimental variable rather than a post-hoc addition.","Language ID tokens, standard in large-scale multilingual ASR, can be counterproductive in bilingual low-resource settings—implying that design choices from high-resource regimes require empirical validation before transfer to low-resource scenarios.","The saturation of multilingual benefit at modest source-language volumes (30 vs 60 hours of Sinhala yielding nearly identical Dhivehi WER) suggests that even small amounts of related-language data may suffice for cross-lingual transfer.","The Turkish control methodology—using an unrelated language at matched data volume—provides a template for distinguishing linguistic relatedness effects from generic data augmentation in other transfer learning studies."],"fun_headline_variants":["Sinhala Transfer Improves Low-Resource Dhivehi ASR","Related-Language Transfer Boosts Dhivehi ASR Accuracy","Continual Pre-Training From Sinhala Enhances Dhivehi ASR","Decoding Config and Sinhala Transfer Drive Dhivehi ASR Gains","Turkish Control Confirms Sinhala Transfer Aids Dhivehi ASR"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The continual pre-training experiment uses a different base model (XLS-R) than all other experiments (Wav2Vec2-BERT), so the paper's claim that CPT is the best transfer strategy is confounded—XLS-R may simply be a stronger architecture independent of the transfer approach.","fun_headline_variants_meta":{"raw":{"variants":["Sinhala Transfer Improves Low-Resource Dhivehi ASR","Related-Language Transfer Boosts Dhivehi ASR Accuracy","Continual Pre-Training From Sinhala Enhances Dhivehi ASR","Decoding Config and Sinhala Transfer Drive Dhivehi ASR Gains","Turkish Control Confirms Sinhala Transfer Aids Dhivehi ASR","Sinhala Cross-Lingual Transfer Cuts Dhivehi ASR Error Rates","Linguistic Relatedness Drives Dhivehi ASR Improvements","Sinhala Pre-Training and KenLM Cut Dhivehi ASR Errors","Cross-Lingual Transfer From Sinhala Improves Dhivehi ASR","Decoding Strategy Strongly Impacts Dhivehi ASR Transfer"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1464,"prompt_tokens":636,"completion_tokens":828,"prompt_tokens_details":null},"tokens_in":636,"tokens_out":828,"duration_ms":22803,"temperature":1.0,"reasoning_tokens":621,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T10:44:21.260926+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If replacing Sinhala with Turkish in the continual pre-training setup still produced WER improvements over the Dhivehi-only baseline comparable to the Sinhala CPT result, the claim that gains stem from linguistic relatedness would be undermined.","supporting_citations":[],"review_version":1}