Evaluating Parkinson's Disease Detection in Anonymized Speech: A Performance and Acoustic Analysis

Carlos Franzreb; Francisco Teixeira; Ben Luks; Sebastian Möller; Alberto Abad

cs.SDMar 8, 2026

Evaluating Parkinson's Disease Detection in Anonymized Speech: A Performance and Acoustic Analysis

Authors: Carlos Franzreb, Francisco Teixeira, Ben Luks, Sebastian Möller, Alberto Abad

Organizations: German Research Center for Artificial Intelligence, Germany · INESC-ID, Portugal · Technical University of Berlin, Germany · Instituto Superior Técnico, University of Lisbon, Portugal

Abstract

Automatic detection of Parkinson's disease (PD) from speech is a promising non-invasive diagnostic tool, but it raises significant privacy concerns. Speaker anonymization mitigates these risks, but it may suppress the pathological information necessary for PD detection. We assess the trade-off between privacy and PD detection for two anonymizers (STT-TTS and kNN-VC) using two Spanish datasets. STT-TTS provides better privacy but severely degrades PD detection by eradicating prosodic information. kNN-VC preserves macro-prosodic features such as duration and F0 contours, achieving F1 scores only 3-7% lower than original baselines, demonstrating that privacy-preserving PD detection is viable when using appropriate anonymization. Finally, an acoustic distortion analysis characterizes specific weaknesses in kNN-VC, offering insights for designing anonymizers that better preserve PD information.

Figures & tables

Dataset	Anon.	PDD-O	PDD-A	WER
GT-S	None	$79\pm 1$	$78\pm 1$	$13$
	kNN-VC r	$47\pm 1$	$80\pm 1$	$18$
	kNN-VC	$54\pm 1$	$76\pm 0$	$21$
	STT-TTS	$49\pm 0$	$53\pm 1$	$12$
GT-M	None	$80\pm 3$	$64\pm 4$	—
	kNN-VC r	$49\pm 3$	$80\pm 1$	—

Table 1: PDD F1-scores for different gender-preserving anonymizers, datasets and tasks. O and A refer to the training data of the PDD (original or anonymized). The PDD-A results of original data are calculated with kNN-VC’s PDD-A. WERs are computed with Whisper-large only for the sentences tasks (S), as there is no ground truth for the monologue tasks (M).

	Anonymizer
	None		kNN-VC		STT-TTS
Dataset	HC	PD	HC	PD	HC	PD
LS	$0.5$	—	$6.9$	—	$35.6$	—
EdAcc	$3.0$	—	$17.9$	—	$43.3$	—
MLS-ES	$0.2$	—	$2.4$	—	$26.0$	—
GT-S	$1.0$	$1.2$	$20.8$	$21.2$	$36.4$	$38.4$

Table 2: Intra-EERs as percentages of the privacy evaluation for original speech and the two anonymizers. The English datasets are evaluated before the MLS-ES fine-tuning.

Figure 1: EMD and MI per feature, for the monologue task of PC-GITA.

Parkinson's Disease Detection from Speech

See all 11 papers

A Benchmark for Early-stage Parkinson's Disease Detection from Speech
May 13, 2026Terry Yi Zhong, Cristian Tejedor-Garcia, Khiet P. Truong +3Parkinson's Disease Detection from Speech Benchmark Design
Centre for Language Studies, Radboud University, Nijmegen, the Netherlands · Center of Expertise for Parkinson and Movement Disorders, Radboud University Medical Center, Nijmegen, the Netherlands
Label-Free Parkinson's Disease Screening from Face and Voice through Mechanistic Interpretability
Aug 10, 2026Jiaheng Su, Yu SunMechanistic Interpretability Multimodal Anomaly Detection
Independent Researcher, Los Angeles, CA, USA · California State Polytechnic University, Pomona, CA, USA
Motor, Cognitive, or Corpus? What Survives Cross-Lingual Transfer in Speech-Based Parkinsons Disease Detection
Aug 13, 2026Serli Kopar, Sam Gijsen, Abner Hernandez +2Cross-Lingual Transfer Speech Processing
Hertie Institute for AI in Brain Health, University of Tübingen, Tübingen, Germany · Tübingen AI Center, University of Tübingen, Tübingen, Germany · Charité–Universitätsmedizin, Department of Psychiatry and Psychotherapy, Berlin, Germany +1

May 13, 2026J/K move · Enter open · S save

A Benchmark for Early-stage Parkinson's Disease Detection from Speech

Terry Yi Zhong, Cristian Tejedor-Garcia, Khiet P. Truong, Janna Maas+2

Centre for Language Studies, Radboud University, Nijmegen, the Netherlands · Center of Expertise for Parkinson and Movement Disorders, Radboud University Medical Center, Nijmegen, the Netherlands

Parkinson's Disease Detection from Speech Benchmark Design Speech Processing Parkinson's Disease Parkinson's Disease Classification

Open paper AskPDF arXiv

Early-stage Parkinson's disease (EarlyPD) detection from speech is clinically meaningful yet underexplored, and published results are hard to compare because studies differ in datasets, languages, tasks, evaluation protocols, and EarlyPD definitions. To address this issue, we propose the first benchmark for speech-based EarlyPD detection, with a speaker-independent split designed for fair and replicable cross-method evaluation on researcher-accessible datasets. The benchmark covers three common speech tasks and evaluates methods under different training-resource settings. We also present multi-dimensional evaluation breakdowns by dataset, aggregation level, gender, and disease stage to support fine-grained comparisons and clinical adoption. Our results provide a replicable reference and actionable insights, encouraging the adoption of this publicly available benchmark to advance robust and clinically meaningful EarlyPD detection from speech.

Abstract

Parkinson's Disease Detection from Speech

Explore similar work

A Benchmark for Early-stage Parkinson's Disease Detection from Speech

Label-Free Parkinson's Disease Screening from Face and Voice through Mechanistic Interpretability

Motor, Cognitive, or Corpus? What Survives Cross-Lingual Transfer in Speech-Based Parkinsons Disease Detection