The effect of speech pathology on automatic speaker verification: a large-scale study.

Area of Science:

• Speech Processing
• Biometrics
• Data Privacy

Background:

• Accessing reliable pathological speech data is crucial for data-driven speech processing.
• Public datasets risk patient health information exposure through re-identification attacks.

Purpose of the Study:

• To assess the privacy risks associated with pathological speech data using automatic speaker verification (ASV).
• To evaluate the impact of various speech disorders and demographics on re-identification risks.

Main Methods:

• Utilized a deep-learning-driven automatic speaker verification (ASV) approach on a large, real-world pathological speech corpus (>3800 subjects).
• Analyzed re-identification risks across different age groups, speech disorders (dysphonia, dysarthria), and pediatric conditions (cleft lip and palate).

Main Results:

• Pathological speech presents higher privacy breach risks than healthy speech.
• Adults with dysphonia face elevated re-identification risks; dysarthria risks are comparable to healthy speakers.
• Speech intelligibility did not affect ASV performance, but recording environment was critical for pediatric cases.

Conclusions:

• Merging diverse pathological speech data significantly decreases equal error rate (EER), enhancing ASV effectiveness.
• Findings highlight the critical need for robust privacy measures in handling pathological speech data for speaker verification.
• This research informs strategies for safeguarding patient confidentiality in digital healthcare.