🧬 biology

Speech Foundation Models for Parkinson’s Disease Detection: A Layer-Wise Comparison with Handcrafted Acoustics Across Two Cohorts

This study demonstrates that speech foundation models, particularly when utilizing intermediate-to-late encoder layers and soft-voting aggregation, outperform handcrafted acoustic features in detecting Parkinson's disease from sentence reading across two independent cohorts, though they still exhibit demographic performance disparities and show more variable results on sustained vowels.

Original authors: Hadi Sedigh Malekroodi, Byeong-il Lee, Myunggi Yi

Published 2026-08-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hadi Sedigh Malekroodi, Byeong-il Lee, Myunggi Yi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Parkinson's disease is a progressive condition that affects how the brain controls movement, often leading to tremors, stiffness, and slowed motion. While these physical signs are what doctors typically look for, the disease also quietly reshapes the voice long before a patient might stumble or shake. The muscles that control breathing, vocal cords, and the tongue become less coordinated, making speech sound softer, breathier, or strangely flat. Because these changes can appear years before a formal diagnosis, scientists have long hoped to use the voice as a simple, non-invasive way to screen for the disease. For decades, researchers have tried to do this by manually measuring specific quirks in the sound, such as how much the pitch wavers or how steady the volume remains. These handcrafted measurements have worked well, but they are limited because they treat speech like a static snapshot rather than a flowing river of sound.

In recent years, a new type of artificial intelligence has emerged that learns to understand speech by listening to millions of hours of human conversation. These systems, known as foundation models, are trained to recognize words and sentences, but they also build a deep internal map of how sound works. A team of researchers at Pukyong National University in South Korea wondered if these powerful, pre-trained systems could spot the subtle signs of Parkinson's better than the traditional manual measurements. They set out to test whether the AI's internal understanding of speech could serve as a superior tool for detecting the disease, and if so, which parts of the AI's brain were doing the heavy lifting.

The researchers tested their ideas using voice recordings from two distinct groups of Spanish speakers: one group from Colombia and another from Spain. These groups included people diagnosed with Parkinson's and healthy individuals of similar ages and genders. To see how the AI performed under different conditions, they asked the participants to do two specific things. First, they asked them to hold a single vowel sound, like "ah," for as long as possible. This task isolates the voice box and tests the stability of the vocal cords. Second, they asked the participants to read aloud a series of sentences that were carefully chosen to cover a wide range of sounds and rhythms. This task requires the complex coordination of the tongue, lips, and breath, mimicking the flow of normal conversation.

The team fed these recordings into five different families of advanced speech AI models. These models included well-known systems like Whisper and Wav2Vec, as well as newer, larger models like Nemotron and Qwen. Instead of just using the final answer the AI gave, the researchers peered inside the models to see what the AI "thought" at every single step of its processing. They extracted the mathematical representation of the sound from the very first layer, where the AI hears raw noise, all the way to the final layers, where it understands complex meaning. They then compared how well these different layers could distinguish between a person with Parkinson's and a healthy person, pitting the AI's insights against the traditional manual measurements.

The results showed that the AI models were generally better at spotting the disease than the traditional manual measurements, but the advantage depended heavily on what the person was saying. When participants read sentences, the AI models were significantly more accurate, achieving an AUC of about 0.88 in the Colombian group and even higher in the Spanish group. The AI excelled here because reading sentences involves the rapid, coordinated movement of the mouth and breath, which is exactly where Parkinson's causes the most trouble. The traditional measurements, which focus on steady sounds, missed many of these dynamic errors. However, when the participants simply held a vowel sound, the gap narrowed. In some cases, the old-fashioned manual measurements were just as good as the AI, suggesting that for simple, steady sounds, the traditional approach still holds its own.

The researchers also discovered that the AI did not need to use its deepest, most complex layers to find the disease. In fact, the middle layers of the models were often the most effective. This finding makes sense because the deepest layers of these AI models are trained to understand grammar and vocabulary, while the disease affects the physical mechanics of speech. The middle layers seem to strike the perfect balance, capturing the acoustic details and the rhythm of the voice without getting distracted by the meaning of the words. Interestingly, making the AI models larger did not automatically make them better at this specific medical task. A massive model with billions of parameters was not consistently superior to a smaller one, indicating that for detecting Parkinson's, the quality of the specific features matters more than the sheer size of the system.

The study also revealed that the AI was not equally good at spotting the disease in everyone. The models tended to perform better for men than for women, and in one of the groups, they were better at identifying the disease in older adults than in younger ones. This suggests that while the technology is powerful, it does not yet treat all demographics fairly. The researchers noted that these differences likely reflect how the disease manifests differently across groups or how the training data was composed, rather than a flaw in the AI itself. They also found that the AI made more mistakes when decoding the speech of people with Parkinson's, a finding that aligns with the idea that the disease makes speech harder for any system to understand, not just a medical classifier.

Ultimately, the study confirms that these advanced speech models are promising tools for detecting Parkinson's, particularly when analyzing connected speech like reading sentences. They offer a way to capture the complex, flowing nature of the disease that older methods miss. However, the researchers caution that these tools are not yet a perfect replacement for clinical judgment. The technology works best when it listens to the whole story of a person's voice rather than a single sound, and it still needs to be tested on much larger and more diverse groups of people to ensure it works for everyone. The path forward involves refining these models to be fairer across different ages and genders, and validating them in real-world settings where background noise and varying recording conditions might challenge their accuracy.

Technical Summary: Speech Foundation Models for Parkinson's Disease Detection

Problem Statement
Parkinson's disease (PD) manifests early in speech through motor deficits affecting respiration, phonation, and articulation, leading to reduced loudness, breathiness, and imprecise articulation. While automated speech analysis offers a low-cost screening candidate, traditional methods rely on handcrafted acoustic biomarkers (e.g., GeMAPS/eGeMAPS) that quantify static spectral and perturbation features. These features often fail to capture the dynamic articulatory impairments and prosodic variations inherent in continuous speech. Conversely, while deep learning models have advanced speech processing, their application to PD detection has been fragmented. Existing studies typically focus on single architectures (often Wav2Vec 2.0 variants), specific feature layers, or limited cohorts. Furthermore, the rapidly emerging generation of speech foundation models (ASR systems built on large language model backbones or trained at massive multilingual scales, such as Nemotron, Qwen3, and OmniASR) remains unevaluated for pathological speech detection. It is also unclear whether the deepest, lexically specialized layers of these models are optimal for detecting primarily acoustic and prosodic disease markers, or if intermediate layers are more discriminative.

Methodology
The study proposes a unified, speaker-independent pipeline to evaluate layer-wise representations from five speech foundation model families against a handcrafted acoustic baseline (eGeMAPS).

Datasets: Two independent Spanish cohorts were used: PC-GITA (Colombian Spanish, 100 speakers) and NeuroVoz (Castilian Spanish, 107 speakers).
Tasks: Two speech production tasks were analyzed:
1. Sustained Vowels: Isolated phonation to probe vocal fold mechanics (jitter/shimmer).
2. Sentence Reading: Phonetically balanced passages to probe dynamic articulatory and prosodic deficits.
Foundation Models: Five families were evaluated across various parameter scales:
- Whisper (Small, Medium, Large)
- Wav2Vec 2.0 (XLS-R)
- OmniASR (Small, Medium)
- NVIDIA Nemotron 3.5 ASR (0.6B)
- Qwen3-ASR (0.6B, 1.7B)
Feature Extraction: Layer-wise representations were extracted from every encoder layer. Frame-level embeddings were mean-pooled to create fixed-length utterance vectors.
Preprocessing: A pipeline of Pearson correlation filtering ( $\theta=0.90$ ), standardization, and Principal Component Analysis (PCA) was applied to reduce dimensionality before classification.
Classification: Four downstream classifiers (SVM, Logistic Regression, Random Forest, XGBoost) were tested under a 5-fold stratified group cross-validation scheme.
Evaluation Metrics: Performance was measured at the segment level (utterance) and aggregated to the participant level using soft voting. Metrics included Accuracy, F1-score, and ROC-AUC. Additionally, ASR decoding errors (Word Error Rate and Character Error Rate) were computed as classifier-independent proxies for intelligibility.

Key Contributions

Comprehensive Model Comparison: The study evaluates five distinct foundation model families, including recent large-scale ASR systems (Nemotron, Qwen3, OmniASR), against a standardized handcrafted baseline across two cohorts and two speech tasks.
Layer-Wise Profiling: It systematically maps classification performance across encoder depths, challenging the assumption that the deepest layers are always optimal for pathological speech detection.
Demographic Equity Analysis: The paper investigates whether foundation models inherently generalize equitably across gender and age subgroups, revealing cohort-dependent performance disparities.
Task-Dependent Utility: It quantifies the differential utility of learned representations for continuous speech versus sustained phonation.

Results

Sentence Reading: Foundation models consistently outperformed the eGeMAPS baseline. The best configurations achieved segment-level F1-scores of approximately 0.88 in both cohorts, with subject-level ROC-AUCs reaching up to 0.9811 (NeuroVoz) and 0.9352 (PC-GITA). The advantage over the baseline was most pronounced in this task.
Sustained Vowels: The advantage of foundation models was smaller and more cohort-dependent. In PC-GITA, the gap between the best foundation model and eGeMAPS was narrow (ROC-AUC ~0.91 vs. 0.87). In NeuroVoz, the gap was larger (0.90 vs. 0.73), but handcrafted features remained competitive.
Model Scaling: Increasing model size within a family (e.g., Whisper Small vs. Large) did not reliably improve detection performance; larger variants sometimes underperformed smaller ones.
Layer Depth: Intermediate-to-late encoder layers generally yielded the strongest representations. The deepest, most lexically specialized layers were not consistently the most discriminative, supporting the hypothesis that PD markers are primarily acoustic/prosodic rather than lexical.
ASR Decoding Errors: PD utterances exhibited systematically higher Word and Character Error Rates than healthy controls across all tested ASR backends, independent of the downstream classifier.
Demographic Disparities: Foundation models did not eliminate performance gaps. In PC-GITA, models performed significantly better for male speakers. In NeuroVoz, performance favored senior speakers (≥65). Soft voting aggregation improved participant-level discrimination but did not resolve these demographic biases.

Significance and Claims
The paper claims that speech foundation models are promising feature extractors for PD detection, particularly for tasks involving continuous speech where dynamic articulatory and prosodic cues are critical. However, the authors modestly assert that these models are not uniformly superior replacements for handcrafted features, especially for sustained phonation tasks where traditional acoustic descriptors remain competitive.

The study highlights that the "black box" nature of foundation models does not guarantee demographic fairness; performance disparities based on gender and age persist and are cohort-dependent. Consequently, the authors conclude that while foundation models offer a powerful new paradigm, their clinical utility requires validation in larger, demographically balanced, and multilingual cohorts, and that equitable performance cannot be assumed solely from pre-training scale. The work motivates future research into stratified training and calibration to address observed demographic gaps.

Technical Summary: Speech Foundation Models for Parkinson's Disease Detection

More like this