Resolve whether alternative NeMo adapter recipes can prevent regression

Determine whether deeper LoRA target sets, higher-rank LoHA or DoRA configurations, or other multi-LoRA recipes can recover the dysarthric automatic speech recognition regression observed when adapting the Parakeet-TDT-0.6B-v3 and Canary-1B-v2 NeMo backbones.

Background

Simple LoRA adaptation caused both tested NeMo backbones to regress relative to their zero-shot performance: Parakeet degraded on the speaker's dysarthric test set, while Canary collapsed during generation. The experiments did not tune adapter depth, rank, or target modules separately for these architectures.

The authors explicitly state that the observed failure may be recipe-dependent rather than structural. They leave unresolved whether deeper target coverage, higher-rank LoHA or DoRA, or other adapter configurations could restore useful adaptation.

References

We deliberately do not claim the failure to be structural in a stronger sense: we have not run deeper LoRA targets (e.g. the RNN-T joint network on Parakeet, cross-attention-only on Canary) or higher-rank LoHA / DoRA variants on these backbones, and we cannot rule out that one of those recipes would recover the regression.

— Choosing a PEFT Variant for Per-Patient Dysarthric ASR: A Single-Speaker Case Study on Two ASR Bases  (2609.02735 - Muller et al., 2 Sep 2026) in Section 4.3, “NeMo-backbone negative result”; Section Conclusion and future work