Everything a voice carries that is not the words, and how much of it models actually use

Speech models detect Parkinson's disease, depression and dementia at accuracies that look clinically useful. A good deal of that accuracy is not the disease. Patients and controls are often recorded in different rooms, on different equipment, on different days, and a classifier given nothing but the acoustic properties of the recording can score close to a model that was supposed to be hearing pathology.

Work here is about telling those two apart. Three questions run through it.

What can the recording alone achieve? Every clinical number should have to clear the score reachable from recording properties with no speech content in the path. Reported next to the result, not left implicit.

Do models answer from the voice, or from a shortcut? Audio language models transcribe internally, so asking one whether a voice sounds ill and supplying a transcript alongside it measures something other than listening. Asking without any text, and ablating the audio instead, measures what was intended.

Which internal features carry the condition? A feature that belongs to the microphone cannot reappear in a corpus recorded somewhere else. A feature that belongs to the pathology can. That difference turns out to be testable when within-corpus statistics are not.

Speech as a health biomarker. Frozen-encoder probing evaluated across corpora rather than within them, a behavioural study of audio language models asked without supplied text, and sparse-autoencoder interpretability with speaker-clustered intervals throughout. Seven corpora, 603 speakers, four languages. Private while the paper is in progress.

Nima Kelidari, Minoo Ahmadi and Chaitanya Parwatkar, University of Southern California, working with Prof. Mohammad Soleymani.

Contact: kelidari.nima@gmail.com