The research behind A'ber.

A'ber is built on machine learning over published dysarthria corpora and on the established motor learning literature for rehabilitation of motor speech disorders. The research classifier (a research prototype that predicts the corpora's perceptual rating labels; it is not a clinical tool) is trained and evaluated on the UA-Speech and TORGO corpora with centroid speaker normalization preceding Cleanlab supervised label cleaning. Its pooled leave-one-speaker-out quadratic weighted kappa over 23 speakers is 0.750, with a 95% confidence interval of [0.512, 0.893]; across ten seeds the mean is 0.759 with a standard deviation of 0.061. Broken down by class, accuracy is 0.798 for mild and 0.915 for severe but 0.250 for moderate, below the 33% chance rate for a three class problem, on every seed. There has been no evaluation on any corpus outside these two, so generalization beyond them is unmeasured. A'ber was verified on synthetic and public research recordings; it is not clinically validated.

The research classifier

The classifier consumes WavLM features over the acoustic battery and emits per dimension rating estimates rather than a single collapsed score. In the product it only sets a starting practice level and is shown to nobody as a rating. The pipeline enforces speaker normalization before Cleanlab so label cleaning operates on speaker invariant features, and the Cleanlab cap is computed as a percentage of flagged samples, not of total class size, which prevents the historical regression where over confident clean labels squeeze out genuine borderline cases.

The adaptive practice engine

Exercise selection is rule based. Deterministic rules filter the candidate set against guards (audio quality gate, session one validity, device baseline) and a staircase on raw practice measures sets difficulty. Research estimators (Kalman tracking, Elo difficulty) are computed and logged in shadow only; they do not decide what a patient practises and are never shown to a patient.

Exercise design and motor learning

Every exercise is grounded in the motor learning literature on speech rehabilitation. Feedback frequency calibration follows the Schmidt and Lee framework, transitioning from knowledge of performance (rich, real time feedback while the patient is learning the movement) to knowledge of results (faded summary feedback once the patient is consolidating). Cadence differs by sub track because the literature shows respiration and articulation respond to different feedback densities, and the engine honors those differences instead of applying one cadence across the board.

Datasets and ethical use

Model development uses the UA-Speech corpus from the University of Illinois at Urbana-Champaign and the TORGO corpus from the University of Toronto, both publicly released for dysarthria research with appropriate consent. UA-Speech is primarily word and command level so it cannot yield connected speech metrics such as maximum phonation time or articulation rate; TORGO contributes the sustained vowel and connected sentence material that completes the eight dimension coverage. The two corpora are never pooled within a single metric. The IEEE SLT manuscript covering the full methodology is in preparation.