Sequential Adapter Stacking for Cross-Lingual Low-Resource ASR
Thai Thi Thanh Thao Dang, Mengjie Qian, Kate Knill
cs.CL
Quick verdict
Stacking a frozen related-language adapter under a target adapter beats full fine-tuning by 5-8% WER on three unseen languages; controlled, single-backbone.
Summary
The paper tackles adapting large multilingual ASR models to languages absent from Whisper's pre-training. It compares three transfer strategies: warm initialization of a target adapter from a source adapter, attention-based target-inclusive fusion (a lightweight SimAdapter variant), and the proposed Sequential Adapter Stacking (SeqStack), which freezes a monolingually trained source-language adapter (~120h) and stacks a near-zero-initialized trainable target adapter on top. Candidate source languages are ranked by three complementary measures: Whisper encoder representation similarity, BPE token coverage decomposed by script class, and genealogical distance from Glottolog/URIEL. On FLEURS data for Asturian, Assamese, and Xhosa, spanning three tiers of related-language pre-training coverage, SeqStack with the closest source (Spanish, Bengali, Zulu) significantly outperforms full fine-tuning at p<0.05 (MAPSSWE), yielding 5-8% relative WER reductions (e.g., Xhosa 38.5 vs 41.9). A capacity ablation (doubled adapter dimension: 45.6 vs 38.5 on Xhosa) shows the gain is not from extra parameters. In the one-hour regime, SeqStack keeps large advantages on Assamese (47.0 vs 53.6) and Xhosa (47.5 vs 63.9), though full fine-tuning slightly wins on Asturian (21.3 vs 21.8). Main limitations: a single backbone (Whisper medium), only three target languages, no seed variance reported, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SeqStack with the closest related source significantly outperforms full fine-tuning on all three targets with 5-8% relative WER reductions (14.9/34.1/38.5 vs 15.7/36.3/41.9).
- Unlike warm initialization and attention-based fusion, SeqStack never falls below target-only adapter tuning across the nine source-target conditions tested.
- The three-measure similarity framework reliably identifies the best source as an ordinal ranking tool, though it does not predict transfer magnitude.
- With one hour of target data, SeqStack retains large advantages on Assamese (47.0 vs 53.6) and Xhosa (47.5 vs 63.9), while on Asturian full fine-tuning is slightly better.
- The Xhosa-Zulu improvement stems from language-specific transfer rather than added capacity, per the d=512 adapter ablation (45.6 vs 38.5).
Strengths
Matched-condition experimental design
All transfer methods use parameter-matched Pfeiffer adapters (d=256) at every encoder/decoder FFN output, source adapters trained on the same ~120h volume, grid-searched learning rates, early stopping on validation WER, and identical greedy decoding - making Table 3 comparisons internally fair.
Capacity confound explicitly ruled out
Section 4 doubles the baseline adapter dimension (d=512: 45.6% WER on Xhosa) versus SeqStack at d=256 (38.5%), directly testing the more-parameters explanation that most adapter papers leave unaddressed.
Significance testing on headline claims
Honest boundary reporting
Triangulated source selection recipe
Weaknesses
No seed variance or multiple-comparison control
Single backbone and single adapter placement
Weakened fusion baseline
Language-token proxy confound
Abstract overstates one-hour persistence
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This 4-page ICASSP-style paper (Whisper medium, FLEURS data) asks two coupled questions: how should a pre-trained source-language adapter be exploited when adapting to a language Whisper does not support, and does source-target relatedness predict which source to pick? It compares target-only adapter tuning, warm initialization (adapted from LoRA-Whisper), a lightweight target-inclusive AdapterFusion variant (inspired by SimAdapter), and the proposed Sequential Adapter Stacking (SeqStack), which freezes a monolingually trained source adapter (~120h) and trains a near-zero-initialized target adapter stacked on top. Sources are ranked by three measures - deep-band encoder representation cosine (EncSim), script-aware BPE token coverage (TokCov), and Glottolog/URIEL genealogy (GenSim) - which agree on the rankings Asturian-Spanish, Assamese-Bengali, Xhosa-Zulu (Table 1). Central claims: (1) SeqStack with the closest source beats full fine-tuning significantly (MAPSSWE, p<0.05) on all three targets, by 5-8% relative WER; (2) SeqStack never falls below adapter tuning, unlike warm-init and fusion; (3) gains largely persist at one hour of target data; (4) the Xhosa gain is transfer-driven, not capacity-driven.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Sequential Adapter Stacking (SeqStack)?
A parameter-efficient ASR adaptation method: a source-language adapter, trained monolingually on about 120 hours of a related language, is frozen, and a near-zero-initialized target-language adapter is stacked on top and trained only on the small target dataset. On Whisper medium it cuts word error rate by 5-8% relative versus full fine-tuning on three languages Whisper does not support.
Which languages were tested and how large were the gains?
Asturian (source: Spanish), Assamese (source: Bengali), and Xhosa (source: Zulu) from the FLEURS corpus, using Whisper medium. Test WER fell from 15.7/36.3/41.9 (full fine-tuning) to 14.9/34.1/38.5 with SeqStack - 5-8% relative reductions, statistically significant at p<0.05 under the MAPSSWE test.
Does the method still work with only one hour of training data?
Mostly, but not everywhere. With one hour of target data, SeqStack beat full fine-tuning by 12% relative on Assamese (47.0 vs 53.6) and 26% on Xhosa (47.5 vs 63.9). On Asturian, however, full fine-tuning was slightly better (21.3 vs 21.8), so the one-hour advantage is not universal.
Is code or data available for this paper?
The paper uses public datasets (FLEURS, Common Voice, IndicVoices, Swivuriso) and reports hyperparameters in detail, but no code release is mentioned, so exact replication currently requires reimplementation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (MAD-X vs SeqStack wiring) cannot be inspected from extracted text; the architecture assessment relies on the equations and prose.
- This is an arXiv v1 preprint in 4-page conference format and has not undergone journal-style peer review.
- Reference [10] shows a typo (Whispe), and table formatting is flattened in extraction; the paper body otherwise appears complete.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.