Training One Speech Model per Disease Beats Pooling Them Together
Bernard Muller studies what happens to a voice when the nervous system behind it fails. He also lives it: ALS has taken his own speech, and he works through an eye-gaze interface, as the paper declares in its own author statement. When a computer learns to grade how badly a disease has damaged someone's speech, does it matter which disease did the damage? The field has mostly assumed not.
Severity grading is still done by ear, by a clinician. It carries real weight: whether a treatment is working, how fast a disease is moving, how a communication device should be tuned. Automating it needs graded recordings of impaired speech, and those are scarce in every language. So the field's instinct has been to pool. Dysarthria, the umbrella word for speech damaged by nerve injury, gets treated as one thing. A single model is trained on whatever is available, as Google's SpICE was on hundreds of thousands of English recordings.
Muller works at The Scott-Morgan Foundation in Torquay, England. With Antonio Armando Ortiz Barrañón at Tecnológico de Monterrey in Mexico and LaVonne Roberts, who works at the foundation and at SMF Labs in Paris, he tested that assumption directly. Their paper, accepted for presentation at the IEEE Spoken Language Technology Workshop in Palermo in December and posted to arXiv on Sept. 18, 2026, is a controlled comparison rather than a product announcement.
They trained four models on the same publicly available speech model, HuBERT, with the same three-stage recipe, the same catalog of corpora and the same held-out test. One saw every condition at once. The other three saw only cerebral palsy, only Parkinson's, or only ALS. Nothing else differed.
On each condition's held-out speakers, the specialist models won. Macro F1, a score that counts every severity grade equally, rose 22.6% for cerebral palsy, 40.0% for Parkinson's and 32.3% for ALS.
That comparison is internal. The pooled model is one the same three authors built for this test, matched to the specialists in architecture and training. It is not SpICE, not a deployed clinical system, and not the published state of the art. Against SALR, the one external published benchmark the paper runs, the verdict splits: the new model is more accurate and scores lower on macro F1, 0.539 against 0.593.
The Parkinson's figure is the showiest and the least robust. Each model was scored on its own filtered pool, so the authors re-ran the comparison on the 65 speakers the two pools shared. There the Parkinson's advantage collapsed to +0.007, a tie. It recovers to +0.067 under a different readout, and the authors blame a subset that happened to hold no moderate or severe Parkinson's speakers. Parkinson's is also the condition the specialist models handle worst outright, so the biggest improvement and the weakest score are the same cell.
Why pooling should hurt is textbook clinical neurology, not this paper's speculation. Parkinson's produces hypokinetic dysarthria, which leaves the contrasts between consonants largely intact until late in the disease. Cerebral palsy produces spastic dysarthria, in which those same contrasts fall apart early. The two signatures run close to opposite. The clinical rulers do not line up either: an intelligibility percentage for cerebral palsy, the Unified Parkinson's Disease Rating Scale, the ALS Functional Rating Scale. One model trained on all of them is being asked to reconcile scales that measure different things.
The gain could still have been a matter of the final decision layer rather than the representation. The authors tested that, bolting a per-condition classifier onto the pooled model's own representation. The scores barely moved, staying within 0.01 of the pooled baseline and far below the specialist models. The loss is in the representation, not in the classifier reading it.
The specialist models are trainable at all partly because of a labeling method from the team's two earlier papers. It measures a speaker against healthy speakers of the same language along a set of sound contrasts, needs no impaired speech to learn from, and sorts the result into four grades. That produced pseudo-labels, severity grades assigned by software, for 1,181 speakers no clinician had graded. For cerebral palsy, the one condition with enough clinically graded data to run the control, those extra speakers lifted macro F1 by 4.3 points. The same control cannot be run for Parkinson's or ALS, and the authors say so. That leaves the 4.3-point gain specific to cerebral palsy alone.
None of this is a clinical instrument, and the authors say so themselves. Checked against 415 Parkinson's speakers whose clinical grades came from outside the pipeline, the software labels agreed at a pooled Cohen's kappa of 0.069, where 1 is perfect agreement and 0 is chance. The harmonized severity scale, the paper says, should not be read as calibrated from one corpus to the next. And when the cerebral palsy model meets a recording collection it has never seen, the score swings hard: the gap between its best and worst corpus, 0.503, is larger than the whole per-condition effect the paper is about. Recording style dominates. An operational system would also need what this work does not supply: a diagnosis, or a classifier ahead of it, to decide which model a patient belongs to.
Each model was trained once, with no repeat runs at different random seeds. The code, training manifests and model checkpoints are promised on acceptance and are not out, so nobody outside the team can rerun any of it. Most of the Parkinson's and ALS test material comes from one English collection.
The practical implication is narrower than a clinical grader: when different diseases produce different speech patterns, pooling the conditions may cost more than the extra data are worth. Severity, at least as these corpora record it, is not one thing.