Training-Free Cross-Lingual Dysarthria Severity Assessment via Phonological Subspace Analysis in Self-Supervised Speech Representations

Bernard Muller; Antonio Armando Ortiz Barrañón; LaVonne Roberts

cs.CLApr 11, 2026

Training-Free Cross-Lingual Dysarthria Severity Assessment via Phonological Subspace Analysis in Self-Supervised Speech Representations

Authors: Bernard Muller, Antonio Armando Ortiz Barrañón, LaVonne Roberts

Organizations: The Scott-Morgan Foundation, Torquay, United Kingdom · Tecnológico de Monterrey, Monterrey, Mexico

Abstract

Dysarthric speech severity assessment typically requires trained clinicians or supervised models built from labelled pathological speech, limiting scalability across languages and clinical settings. We present a training-free method that quantifies dysarthria severity by measuring degradation in phonological feature subspaces within frozen HuBERT representations. No supervised severity model is trained; feature directions are estimated from healthy control speech using a pretrained forced aligner. For each speaker, we extract phone-level embeddings via Montreal Forced Aligner, compute d-prime scores along phonological contrast directions (nasality, voicing, stridency, sonorance, manner, and four vowel features) derived exclusively from healthy controls, and construct a 12-dimensional phonological profile.Evaluating 890 speakers across 10 corpora, 5 languages (English, Spanish, Dutch, Mandarin, French), and 3 primary aetiologies (Parkinson's disease, cerebral palsy, ALS), we find that all five consonant d-prime features correlate significantly with clinical severity (random-effects meta-analysis rho = -0.50 to -0.56, p < 2e-4; pooled Spearman rho = -0.47 to -0.55 with bootstrap 95% CIs not crossing zero). The effect replicates within individual corpora, survives FDR correction, and remains robust to leave-one-corpus-out removal and alignment quality controls. Nasality d-prime decreases monotonically from control to severe in 6 of 7 severity-graded corpora. Mann-Whitney U tests confirm that all 12 features distinguish controls from severely dysarthric speakers (p < 0.001).The method requires no dysarthric training data and applies to any language with an existing MFA acoustic model (currently 29 languages). We release the full pipeline and phone feature configurations for six languages.

Figures & tables

Figure 1: Phonological subspace collapse in dysarthric speech. UMAP projection of phone-level HuBERT embeddings for one healthy control speaker (MC01, left) and one speaker with severe dysarthria (M01, right) from the TORGO corpus. Points represent individual phone tokens coloured by class: nasal consonants (m, n, ng) vs. oral stops (p, b, t, d, k, g). In the healthy speaker, the two classes form tight, well-separated clusters ( $d^{\prime}=4.2$ ). In the severe speaker, the class centroids remain distinguishable but the within-class distributions are wider and overlap substantially ( $d^{\prime}=1.1$ ), reflecting degraded articulatory contrast rather than complete phonological merger. This reduction in separability — quantified as $d^{\prime}$ — is the basis of the severity metric developed in this paper. Note that $d^{\prime}$ is computed along the optimal discriminant direction in 768-dimensional embedding space, not in the 2D UMAP projection shown here.

Figure 2: Clinical validation: $d^{\prime}$ correlates with clinician-rated intelligibility across languages. Left: stridency $d^{\prime}$ vs intelligibility (Kendall’s $\tau=0.407$ , $p=2.9\times 10^{-6}$ , $n=64$ ). Right: mean consonant $d^{\prime}$ vs intelligibility ( $\tau=0.323$ , $p=3.9\times 10^{-5}$ , $n=86$ ). Data from UA-Speech (English, cerebral palsy) and COPAS (Dutch, mixed aetiologies). Trendlines show positive linear association.

Feature	Positive class ( $P_{k}$ )	Negative class ( $N_{k}$ )	Articulatory basis
Nasality	/m, n, ng/	/p, b, t, d, k, g/	Velopharyngeal port opening
Voicing	/b, d, g, v, z, …/	/p, t, k, f, s, …/	Vocal fold vibration
Stridency	/s, z, sh, zh, f, v, …/	/p, t, k, b, d, g, m, n, l, r/	Turbulent airflow at constriction
Sonorance	/m, n, ng, l, r, j, w/	/p, b, t, d, k, g, f, v, s, z, …/	Periodic voicing without obstruction
Manner	/p, b, t, d, k, g, m, n, ng/	/f, v, s, z, sh, zh, …/	Complete vs. partial oral closure

Consonant features (5).

Feature	Positive class	Negative class	Articulatory basis
High	/i, I, u, U/	/A, ae, E, O, @, V/	Tongue height (close)
Low	/A, ae/	/i, I, u, U, @, E/	Tongue height (open)
Back	/u, U, O, A, V/	/i, I, E, ae/	Tongue backness
Round	/u, U, O/	/i, I, E, A, ae, @, V/	Lip rounding

Vowel features (4).

Intelligibility	Severity label
$>94\%$	Control
85–94%	Mild
70–84%	Moderate
$<70\%$	Severe

Intelligibility-based severity thresholds (Stipancic et al. [ 29 ] ).

Corpus	Language	Aetiology	Control	Mild	Moderate	Severe	Total	Source
SAP	English	PD, ALS, CP, DS, Stroke	0*	159	24	5	188	Millet et al. [ 30 ] Zheng et al. [ 31 ]
COPAS	Dutch	Mixed pathology	44	74	77	23	218	Martens et al. [ 32 ]
TORGO	English	CP, ALS	7	2	2	4	15	Rudzicz et al. [ 33 ]
UA-Speech	English	CP	13**	5	0	10	28	Kim et al. [ 34 ]
Neurovoz	Spanish	PD	58	43	10	0	111	Moro-Velazquez et al. [ 35 ]
YouTube_French	French	ALS	20	1	3	1***	25***	Collected by authors

Table 1: Datasets used in the full analysis pipeline. All 12 metrics computed via MFA-aligned speech except where noted. * SAP has no matched controls; 150 LibriSpeech speakers serve as the English healthy control reference for feature direction computation. ** UA-Speech controls from the UASPEECH_control partition (13 speakers). *** One speaker recorded at both moderate and severe stages. **** Speakers serving as healthy controls contribute to both direction estimation and severity evaluation (as the control group).

Corpus	Language	Aetiology	Control	Dysarthric	Method	Source
VOC-ALS	Italian	ALS	72	81	Sustained /a/, /i/, /u/	Mulfari et al. [ 39 ]
PC-GITA	Spanish	PD	50	50	Sustained /a/, /i/, /u/	Orozco-Arroyave et al. [ 37 ]

Table 2: Additional datasets for vowel triangle analysis only (sustained vowel recordings, no MFA required).

Metric	Control	Mild	Moderate	Severe	Ctrl-to-Sev change
Nasality $d^{\prime}$	3.55	3.57	3.15	2.55	$-28.2\%$
Voicing $d^{\prime}$	2.75	2.90	2.74	2.14	$-22.2\%$
Manner $d^{\prime}$	2.82	2.51	2.59	2.29	$-18.8\%$
Round $d^{\prime}$	2.12	–	1.66	1.35	$-36.2\%$
Low $d^{\prime}$	1.86	–	1.82	1.87	$+0.8\%$
Boundary sharpness	0.354	0.424	0.487	0.524	$+48.0\%$

Table 3: Proof-of-concept results: severity group means for selected phonological features (YouTube_French,

N=7

speakers, ALS aetiology).

Figure 3: Left panel: group means for all reported features across control/mild/moderate/severe stages (Table 3 ). Right panel: Speaker A (progressing, voicing $d^{\prime}$ drops 0.37 between moderate and severe) vs. Speaker B (stable in early disease). Data from Table 3 .

Figure 4: Consonant $d^{\prime}$ degradation by severity across corpora. Each panel displays one of the five consonant $d^{\prime}$ features (nasality, sonorant, voicing, strident, manner) with speaker-level means plotted across severity groups (control, mild, moderate, severe) for each corpus independently. Lines connect severity group means within each corpus. All five features show a consistent downward trend from control to severe, with the steepest declines in MDSC (Mandarin, CP) and UA-Speech (English, CP). Corpora with fewer than three severity levels are omitted from individual panels.

Corpus (lang, aetiology)	Nasality	Sonorant	Voicing	Strident	Manner
MDSC (zh, CP)	$-0.923$ ***	$-0.910$ ***	$-0.879$ ***	$-0.901$ ***	$-0.912$ ***
UA-Speech (en, CP)	$-0.786$ ***	$-0.733$ ***	$-0.655$ ***	$-0.726$ ***	$-0.691$ ***
TORGO (en, CP/ALS)	$-0.735$ ***	$-0.681$ ***	$-0.659$ ***	$-0.602$ **	$-0.687$ ***
Neurovoz (es, PD)	$-0.514$ ***	$-0.453$ ***	$-0.349$ ***	$-0.412$ ***	$-0.387$ ***
COPAS (nl, mixed)	$-0.478$ ***	$-0.411$ ***	$-0.325$ ***	$-0.389$ ***	$-0.356$ ***
SAP (en, mixed)	$-0.404$ ***	$-0.394$ ***	$-0.374$ ***	$-0.381$ ***	$-0.376$ ***

Table 4: Within-corpus severity correlations demonstrate the effect is not a pooling artifact. Spearman

\rho

for consonant

d^{\prime}

features vs. ordinal severity. Only corpora with at least 3 severity levels shown. *

p<0.05

, **

p<0.01

, ***

p<0.001

Figure 5: Severity correlation forest plot for all 12 phonological features. Spearman $\rho$ between each feature and ordinal severity (control $=0$ , mild $=1$ , moderate $=2$ , severe $=3$ ), pooled across all corpora. Horizontal bars show bootstrap 95% confidence intervals (1,000 iterations). Features are sorted by effect size. Dark navy: consonant $d^{\prime}$ features; light blue: vowel $d^{\prime}$ features; orange: structural metrics. All $d^{\prime}$ features correlate negatively with severity (CIs not crossing zero); boundary sharpness and cross-position cosine similarity show near-zero pooled correlations due to the speech-type confound discussed in the text.

Figure 6: Consonant $d^{\prime}$ profiles across 867 speakers sorted by severity. Each column is one speaker; each row is one consonant $d^{\prime}$ feature. Speakers are sorted left to right by severity (control, mild, moderate, severe), then by mean $d^{\prime}$ within each group. Top strips show severity (green to red) and aetiology (colour-coded). The green-to-red gradient across all five features demonstrates that phonological subspace degradation is consistent, gradual, and present across aetiologies and corpora.

Feature	Severity	English	Spanish	Dutch	Mandarin	French
Nasality $d^{\prime}$	Control	4.31	2.84	3.72	5.64	3.55
	Severe	1.57	1.35*	1.94	1.48	2.55
Strident $d^{\prime}$	Control	2.95	3.09	2.77	3.63	2.25
	Severe	0.99	1.10*	1.21	0.71	1.76
Sonorant $d^{\prime}$	Control	3.65	2.00	1.86	3.99	1.97
	Severe	1.54	0.89*	1.29	0.98	1.70

Table 5: Mean consonant

d^{\prime}

by severity and language. Values are means across speakers within each severity-language cell. * Spanish severe values from PC-GITA (

n=4

severe speakers).

Speech type	$\rho$	$p$	$n$ (speakers)	Direction
Read sentences	$-0.10$ *	0.01	–	Slight blurring
Spontaneous speech	$-0.12$ ns	–	24	Blurring (expected)
Single words	$+0.28$ ns	–	28	Sharpening
Mixed (SAP)	$+0.32$ ***	5e-6	188	Sharpening

Table 6: Boundary sharpness vs. severity stratified by speech type (Spearman

\rho

). *

p<0.05

, ***

p<0.001

, ns = not significant.

Corpus	Content type	$\rho$	Direction
PC-GITA	Read sentences	$+0.46$ ***	Sharpening
MDSC	Read words	$+0.47$ ***	Sharpening
SAP	Mixed tasks	$+0.32$ ***	Sharpening
COPAS	Read sentences	$+0.18$ **	Sharpening
YouTube_French	Spontaneous / connected	$-0.12$ ns	Blurring

Table 7: Boundary sharpness vs. severity by corpus (Spearman

\rho

). **

p<0.01

, ***

p<0.001

, ns = not significant.

Figure 7: Phonological fingerprints by aetiology. Radar plots showing the mean consonant

d^{\prime}

profile for cerebral palsy (CP,

n=86

), Parkinson’s disease (PD,

n=211

), and amyotrophic lateral sclerosis (ALS,

n=50

) speakers. Each axis represents one consonant

d^{\prime}

feature; the green shaded area shows the healthy control reference profile. CP shows the most uniform degradation across all features, while PD and ALS exhibit relatively preserved voicing. These aetiology-specific profiles demonstrate that the method captures clinically meaningful differences in articulatory impairment patterns.

Feature	CP ( $n=86$ )	PD ( $n=211$ )	ALS ( $n=50$ *)
Nasality $d^{\prime}$	$-0.47$	$-0.38$	$-0.365$ **
Sonorant $d^{\prime}$	$-0.42$	$-0.35$	$-0.365$ **
Voicing $d^{\prime}$	$-0.35$	$-0.29$	$-0.28$
Strident $d^{\prime}$	$-0.44$	$-0.33$	$-0.317$ **
Manner $d^{\prime}$	$-0.30$	$-0.31$	$-0.30$
Boundary sharpness	$+0.27$	–	–

Table 8: Per-aetiology Spearman

\rho

for consonant

d^{\prime}

features vs. severity. Only aetiologies with

n\geq 50

speakers and 3+ severity levels shown. * ALS limited to 50 speakers from SAP and YouTube_French, with only 5 severity-rated (YouTube_French); SAP ALS speakers are predominantly mild. **

p<0.01

despite small

n

; strong trend but limited statistical power.

Feature	$\rho$ ( $d^{\prime}$ vs $n_{\text{phones}}$ , HC only)
Sonorant $d^{\prime}$	$+0.554$
Nasality $d^{\prime}$	$+0.405$
Voicing $d^{\prime}$	$+0.380$
Strident $d^{\prime}$	$+0.350$

Table 9: Spearman correlation between

d^{\prime}

and phone token count (

n_{\text{phones}}

) in healthy control speakers only.

Figure 8: Token count versus nasality $d^{\prime}$ across corpora. Each bubble represents one corpus, positioned by median phone tokens per speaker ( $x$ -axis) and mean nasality $d^{\prime}$ in healthy controls only ( $y$ -axis). Bubble size is proportional to the number of speakers; colour indicates language. The positive association between token count and $d^{\prime}$ in healthy controls reveals a confound: corpora with shorter utterances yield lower $d^{\prime}$ estimates regardless of severity, motivating the partial correlation analyses reported in the Token Count Confound section.

Corpus (aetiology)	Control	Mild	Moderate	Severe	Ctrl-to-Sev
VOC-ALS (ALS)	28.79	27.45	22.73	17.31	$-39.9\%$
PC-GITA (PD)	21.21	19.54	17.17	19.12*	$-9.3\%$ **

Table 10: HuBERT vowel triangle area by severity (sustained vowel datasets). * PC-GITA severe cell contains only 4 speakers; the bounce at severe is likely a small-sample artifact. ** Control-to-moderate change.

Aetiology	Recall
PD	53%
CP	76%
ALS	low (limited by $n$ )

Per-aetiology LOSO recall.

Figure 9: ROC curves for binary severity detection. Left: severe vs. rest (stridency $d^{\prime}$ AUC $=0.890$ ). Right: moderate-or-worse vs. mild/control. All five consonant $d^{\prime}$ features and their mean are shown. Optimal thresholds determined by Youden’s J statistic.

Feature	Raw $\rho$	Partial $\rho$ (controlling $n_{\text{phones}}$ )	Change
Nasality $d^{\prime}$	$-0.547$	$-0.548$	$<0.001$ (unchanged)
Strident $d^{\prime}$	$-0.486$	$-0.504$	$+0.018$ (strengthened)
Sonorant $d^{\prime}$	$-0.529$	$-0.527$	$-0.002$ (unchanged)
Voicing $d^{\prime}$	$-0.527$	$-0.525$	$-0.002$ (unchanged)
Manner $d^{\prime}$	$-0.474$	$-0.476$	$+0.002$ (unchanged)

Table 11: Partial Spearman correlations controlling for phone token count. All

p<10^{-48}

Quartile	Token range	Severity-nasality $\rho$
Q1 (lowest)	Fewest phones	$-0.13$
Q2		$-0.37$
Q3		$-0.57$
Q4 (highest)	Most phones	$-0.37$

Token-count quartile stratification (severity-nasality

\rho

Feature	$\rho$	95% CI	CI width
Nasality $d^{\prime}$	$-0.547$	$[-0.597,-0.495]$	0.102
Sonorant $d^{\prime}$	$-0.529$	$[-0.578,-0.468]$	0.110
Voicing $d^{\prime}$	$-0.527$	$[-0.576,-0.475]$	0.101
Strident $d^{\prime}$	$-0.486$	$[-0.541,-0.428]$	0.113
Manner $d^{\prime}$	$-0.474$	$[-0.530,-0.414]$	0.116

Table 12: Bootstrap confidence intervals.

Corpus removed	Remaining $n$	Nasality $\rho$	Change from full
None (full)	829	$-0.547$	–
LibriSpeech	679	$-0.37$	$+0.18$ (weakened)
PC-GITA	729	$-0.66$	$-0.11$ (strengthened)
SAP	641	$-0.55$	$<0.01$
COPAS	602	$-0.56$	$-0.01$
TORGO	814	$-0.55$	$<0.01$

Table 13: Leave-one-corpus-out sensitivity analysis for nasality

d^{\prime}

(

\rho

with ordinal severity). All correlations remain negative.

Figure 10: Random-effects meta-analysis forest plot, $k=8$ corpora. $I^{2}=87$ – $92\%$ indicates high between-corpus heterogeneity in magnitude but consistent direction. All five consonant $d^{\prime}$ features are significant under both DerSimonian-Laird and the more conservative Hartung-Knapp-Sidik-Jonkman estimator (all HKSJ $p<0.013$ ).

Feature	HuBERT-base	WavLM-base	wav2vec2-base
Nasality	$-0.619$ *	$-0.577$ *	$-0.783$ *
Voicing	$-0.669$ *	$-0.663$ *	$-0.682$ *
Strident	$-0.657$ *	$-0.654$ *	$-0.627$ *
Sonorant	$-0.711$ *	$-0.682$ *	$-0.640$ *
Manner	$-0.674$ *	$-0.699$ *	$-0.636$ *

Table 14: Spearman

\rho

across SSL models (TORGO,

n=15

Dysarthria Assessment

See all 11 papers

Apr 23, 2026

Phonological Subspace Collapse Is Aetiology-Specific and Cross-Lingually Stable: Evidence from 3,374 Speakers

We previously introduced a training-free method for dysarthria severity assessment based on d-prime separability of phonological feature subspaces in frozen self-supervised speech representations, validated on 890 speakers across 5 languages with HuBERT-base. Here, we scale the analysis to 3,374 speakers from 25 datasets spanning 12 languages and 5 aetiologies (Parkinson's disease, cerebral palsy, ALS, Down syndrome, and stroke), plus healthy controls, using 6 SSL backbones. We report three findings. First, aetiology-specific degradation profiles are distinguishable at the group level: 10 of 13 features yield large effect sizes (epsilon-squared > 0.14, Holm-corrected p < 0.001), with Parkinson's disease separable from the articulatory execution group at Cohen's d = 0.83; individual-level classification remains limited (22.6% macro F1). Second, profiles show cross-lingual profile-shape stability: cosine similarity of 5-dimensional consonant d-prime profiles exceeds 0.95 across the languages available for each aetiology. Absolute d-prime magnitudes are not cross-lingually calibrated, so the method supports language-independent phenotyping of degradation patterns but requires within-corpus calibration for absolute severity interpretation. Third, the method is architecture-independent: all 6 backbones produce monotonic severity gradients with inter-model agreement exceeding rho = 0.77. Fixed-token d-prime estimation preserves the severity correlation (rho = -0.733 at 200 tokens per class), confirming that the signal is not a token-count artefact. These results support phonological subspace analysis as a robust, training-free framework for aetiology-aware dysarthria characterisation, with evidence of cross-lingual profile-shape stability and cross-backbone robustness in the represented sample.

Bernard Muller, Antonio Armando Ortiz Barrañón, LaVonne Roberts

The Scott-Morgan Foundation, Torquay, United Kingdom · Tecnológico de Monterrey, Monterrey, Mexico

Dysarthria Assessment

Ask PDF

Sep 18, 2026

Per-Aetiology Contrastive Severity Embeddings with Phonological Pseudo-Labelling for Multilingual Dysarthric Speech

Most multilingual dysarthria-severity systems either train on a single aetiology-language pair or pool heterogeneous aetiologies into one label space. We test that pooling assumption with four matched HuBERT-base contrastive embedding models under a shared backbone, training recipe, corpus registry and held-out evaluation: one mixed-aetiology baseline and three aetiology-specific models for cerebral palsy (CP), Parkinson's disease (PD) and amyotrophic lateral sclerosis (ALS). Training combines clinically labelled speech with ordinal pseudo-labels from a training-free phonological profiling method [1], [2]. On speaker-disjoint, leakage-filtered held-out subsets, the per-aetiology models outperform the mixed baseline across all three target aetiologies: CP (macro F1 0.829 vs 0.676, +22.6 % relative), PD (0.715 vs 0.511, +40.0 %) and ALS (0.788 vs 0.596, +32.3 %). On CP, adding 144 SAP and 44 CDSD pseudo-labelled speakers lifts macro F1 from 0.786 to 0.829 over a clinical-only CP model (+4.3 percentage points). Training data span three to seven languages per aetiology. We position this as a controlled comparison of label-space design choices and discuss pseudo-label calibration, split hygiene, and confidence-thresholded deployment as important limitations for future work.

Bernard Muller, Antonio Armando Ortiz Barrañón, LaVonne Roberts

The Scott-Morgan Foundation, Torquay, United Kingdom · Tecnol´ogico de Monterrey, Monterrey, Mexico · SMF Labs, Paris, France

Speech Quality Assessment Dysarthria Assessment

Ask PDF

Jun 22, 2026

Cross-lingual Retrieval-Augmented Classification for Dysarthria Severity Assessment

Automatic dysarthria severity assessment is limited by the scarcity of labeled pathological speech data. To address this, we propose Cross-lingual Retrieval-Augmented Classification (CRAC), which leverages speech from a different language via an align-retrieve-fuse pipeline. Supervised contrastive learning first shapes a severity-focused embedding space, then a vector database is built from the opposite-language corpus. During both training and inference, the classifier retrieves top-k references from the aligned space and fuses them with the input via cross-attention. Evaluated on Korean post-stroke and Italian ALS dysarthria datasets under a speaker-independent three-class protocol, CRAC achieves balanced accuracies of 87.3% on Korean and 86.7% on Italian, improving over monolingual baselines by 8.4 and 20.0 percentage points, respectively.

Taeyoung Jeong, Insung Lee, Du-Seong Chang +1

Department of Artificial Intelligence, Sogang University, South Korea

Retrieval-Augmented Classification Speech Quality Assessment Speech Processing Dysarthria Assessment Clinical Speech Processing Cross-Lingual Transfer

Ask PDF

Abstract

Dysarthria Assessment

Explore similar work

Phonological Subspace Collapse Is Aetiology-Specific and Cross-Lingually Stable: Evidence from 3,374 Speakers

Per-Aetiology Contrastive Severity Embeddings with Phonological Pseudo-Labelling for Multilingual Dysarthric Speech

Cross-lingual Retrieval-Augmented Classification for Dysarthria Severity Assessment