Working papers · speech, health & trustworthy AIVigo, 2026

Research / Voice as a clinical signal / [2]

Radial LoRA: one parameter that makes LoRA work on clinical speech

Radial LoRA: Enhancing Parameter-Efficient Fine-Tuning for Clinical Speech Assessment. IberSPEECH 2026, 2026.

In one sentence

Tempering the size of the adapter's input with one learnable exponent per layer lifts Macro-F1 by 12.4 points over LoRA at the same parameter budget (p = 4.4 × 10⁻¹⁰) across four clinical speech corpora, and beats full fine-tuning with about 322 times fewer trainable parameters.

+12.4Macro-F1 points over LoRA
4clinical corpora
+24parameters over LoRA
< 1%added inference time

The problem

Large self-supervised speech models transfer well, but clinical speech datasets hold a few hundred speakers at most. Full fine-tuning overfits them, and parameter-efficient methods such as LoRA work less well when the information that separates healthy from pathological speech lies in the direction of the hidden representations rather than in their size. LoRA’s update is linear, so it treats both the same way.

What we propose

R-LoRA (Radial Low-Rank Adaptation) passes the adapter’s input through a radial operator before the low-rank projection: it divides the input by its norm raised to a learnable exponent λ between 0 and 1. At λ = 0 it is LoRA; at λ = 1 the update sees only the direction of the input. Each adapted projection learns its own λ, so R-LoRA adds one parameter per projection: 24 on Wav2Vec2 base.

A geometric audit of the frozen model supports the premise. On all four corpora the representations are anisotropic (mean pairwise cosine 0.81 to 0.94), use at most 12 % of the available dimensions, and separate healthy from pathological speakers by small angles.

What we found

Across four clinical corpora, eight ranks and five outer folds, R-LoRA beats LoRA by 12.4 points of Macro-F1 (95 % CI 8.9 to 15.8; p = 4.4 × 10⁻¹⁰). It wins 113 of 160 paired comparisons, is significantly better on three corpora and never significantly worse. Its margin is widest at low rank (+17.4 points at rank 2) and still large at rank 256 (+10.7).

On VisIA, the suicidal-ideation corpus, LoRA and full fine-tuning both collapse to the majority class and detect no case of active ideation; R-LoRA reaches 50 % recall of that class with 80 % precision. Across corpora it also raises recall of the pathological class, which is what matters for screening.

Full fine-tuning trails at every rank, by 14.2 points on average: with about 150 labelled speakers per corpus, updating all 95 million weights overfits.

The learned exponents stay between the two extremes: λ ranges from 0.08 to 0.94 with median 0.52, so the model chooses partial normalisation, never plain LoRA and never direction only.

The trade-off

LoRA’s update can be merged into the weights after training, so it costs nothing at inference. R-LoRA’s cannot, because its operator is non-linear. Measured at rank 8, inference takes 3.30 against 3.28 ms per sample, within run-to-run noise, and training runs at 0.96 times LoRA’s throughput.

Why it matters

R-LoRA is a drop-in change to LoRA that pays off where data are scarce and representations are collapsed, which describes most clinical speech. It is defined on the adapter’s input, not on a particular model. The method is explained step by step in Radial LoRA: when the direction matters more than the size.

0255075100 NeuroVoz 1.1 COPERIA /a/ 14.1 COPERIA cough 10.2 VisIA 24.1 All corpora 12.4
Figure 1. Mean Macro-F1 gain of R-LoRA over LoRA (points), averaged over eight ranks and five outer folds. Significant on all but NeuroVoz.

Cite

Ramírez-Sánchez J.M., Docío-Fernández L., García-Caballero A.A., García-Mateo C. Radial LoRA: Enhancing Parameter-Efficient Fine-Tuning for Clinical Speech Assessment. IberSPEECH 2026, 2026.