Choosing a PEFT Variant for Per-Patient Dysarthric ASR: A Single-Speaker Case Study on Two ASR Bases
Bernard Muller, László Tóth, LaVonne Roberts
cs.CL · cs.SD
Quick verdict
Case study: on one dysarthric speaker, LoRA ties DoRA, beats 4-bit QLoRA and 4 other PEFT variants on two ASR bases; careful work, but n=1 limits generality.
Summary
This single-speaker case study compares seven LoRA-family parameter-efficient fine-tuning (PEFT) variants — LoRA, QLoRA, DoRA, AdaLoRA, LoHA, VeRA and VB-LoRA — for per-patient dysarthric ASR, asking which adapter to default to when each patient trains on their own data. The authors train adapters for one post-stroke Hungarian speaker with severe dysarthria (409 utterances; 262 train) on two bases: Whisper-large-v3 with Hungarian fine-tuning (29.46% zero-shot CER) and a private multilingual Qwen3-ASR-1.7B production checkpoint (49.46%). A three-seed head-to-head with a paired bootstrap finds LoRA and DoRA statistically tied (13.86 vs 13.90% CER on Whisper; 28.10 vs 28.33% on Qwen3; p>0.5), while real 4-bit QLoRA is worse on every seed with no VRAM saving. LoHA, VeRA, VB-LoRA and AdaLoRA fail to approach the LoRA family under the fixed ~80-step budget, though LoHA still cuts Whisper CER by 18.6% relative. Full fine-tuning is more accurate (11.43%), but a 115 MB LoRA extended to feed-forward blocks comes within 0.66 pp at ~3.7% of the storage. An enrollment grid shows ~5 minutes of audio yields 45.6% of the zero-shot-to-30-minute improvement. Main limitations: one speaker, one language, a private second base, and a reused 107-utterance test set; the authors themselves flag all of these.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- In the speaker-dependent per-patient regime (~80 optimizer steps), LoRA and DoRA are statistically tied on both bases (13.86 vs 13.90% CER on Whisper; 28.10 vs 28.33% on Qwen3), so the simpler LoRA is adopted as the production default.
- Real 4-bit NF4 QLoRA is worse than LoRA on every seed on both bases (+0.69 pp CER on Whisper, +1.99 pp on Qwen3) and yields no peak-VRAM saving at 1.5-1.7B scale.
- LoHA, VeRA, VB-LoRA and AdaLoRA do not reach the LoRA family under the fixed small-data budget, though LoHA still delivers an 18.6% relative CER reduction on Whisper.
- A 115 MB LoRA extended to feed-forward blocks reaches 12.09% CER, within 0.66 pp of full fine-tuning (11.43%), at roughly 3.7% of its per-patient storage, with encoder attention the largest single lever (Table 8).
- About 5 minutes of enrollment audio captures 45.6% of the zero-shot-to-30-minute CER reduction for this speaker, with substantial further gains at 10 and 30 minutes (Table 6).
Strengths
Controlled-variable cross-variant design
Same speaker, same data, one fixed recipe; only the PEFT variant varies across the 16-cell Table 1 screen, with per-variant trainable-parameter counts, serialized adapter sizes and wall-clock times reported, so the comparison surface is fully explicit (Section 3.3).
Statistically disciplined headline decision
Section 4.2 runs a 10,000-resample paired bootstrap over pooled per-utterance hypotheses (Whisper delta +0.03 pp, CI [-0.17, +0.25], p=0.79; Qwen3 +0.23 pp, CI [-0.50, +0.96], p=0.55) and explicitly refuses an equivalence claim absent a pre-registered margin, settling for 'practical tie'.
Negative results analyzed, not buried
Attribution ladder localizes the full-FT gap
Unusual transparency
Weaknesses
n=1 evidence base
Fixed-budget screen may produce false negatives
Half the study sits on a private base
Reused evaluation set
Recipe tuned for one adapter class; unresolved reference ambiguity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a production-oriented case study, not a benchmark: one post-stroke Hungarian male speaker (S1; 409 utterances, split 262 train / 40 val / 107 text-disjoint test), seven LoRA-family PEFT variants (LoRA, QLoRA, AdaLoRA, DoRA, LoHA, VeRA, VB-LoRA), two ASR bases (Whisper-large-v3 with a merged 38K-utterance Hungarian fine-tune, zero-shot 29.46% CER; a private multilingual Qwen3-ASR-1.7B production checkpoint, zero-shot 49.46%), and one fixed recipe (AdamW 1e-4, bf16, batch 4 x grad-acc 4, 5 epochs = ~80 optimizer steps, greedy decoding, CER primary). Headline claims: (1) attention-projection adapters work on both bases; (2) LoRA and DoRA are statistically tied, so the simpler LoRA becomes the per-patient default; (3) real 4-bit QLoRA loses on every seed with no VRAM saving; (4) LoHA, VeRA, VB-LoRA and AdaLoRA do not reach the LoRA family under this budget; (5) a 115 MB feed-forward-extended LoRA comes within 0.66 pp CER of full fine-tuning at ~3.7% of its storage; (6) ~5 minutes of enrollment audio captures 45.6% of the zero-shot-to-30-minute CER reduction.
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
Which PEFT method is best for per-patient dysarthric ASR?
According to this single-speaker study, LoRA r=16 is the recommended default: it statistically ties DoRA (13.86 vs 13.90% CER on Whisper), trains faster, and beats real 4-bit QLoRA on every seed, while VeRA, VB-LoRA and AdaLoRA show almost no adaptation within the ~80-step budget. The evidence comes from one severe post-stroke Hungarian speaker, so multi-speaker validation is still pending.
How much audio is needed to personalize ASR for a dysarthric speaker?
In this study, about 5 minutes of patient audio already captured 45.6% of the CER improvement achievable with 30 minutes (dropping from 29.46% zero-shot to 22.49%), with gains continuing at 10 minutes (18.87%) and 30 minutes (14.17%). The curve was measured on a single severe post-stroke speaker, so the shape may differ for other severities and aetiologies.
Is 4-bit QLoRA a good choice for fine-tuning Whisper on small data?
Not in this setting: real 4-bit NF4 QLoRA was worse than plain LoRA on every seed on both bases (14.56 vs 13.86% CER on Whisper; 30.09 vs 28.10% on Qwen3-ASR) and, at 1.5-1.7B parameter scale, actually saved no peak VRAM (16.9 vs 14.8 GiB on Whisper). The authors recommend against 4-bit quantization in this small-data per-patient regime.
Can LoRA match full fine-tuning for dysarthric ASR?
Almost: full fine-tuning of Whisper-large-v3 was more accurate (11.43 vs 13.86% CER), but extending LoRA from attention projections to also cover the feed-forward blocks closed most of the gap (12.09% CER) using a 115 MB adapter — about 3.7% of the per-patient storage of a full 3.1 GB checkpoint.
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.
- The task's metadata abstract was cut mid-sentence ('one speaker, o'), but the complete abstract is present at the start of the extracted full text, so no content was missing.
- Figures 1-2 are available only as captions and partial axis/legend labels in the extraction; all numeric judgments rely on the tables and prose, which are 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.