Addressing challenges in speaker anonymization to maintain utility while ensuring privacy of pathological speech

Area of Science:

• Speech processing
• Biomedical informatics
• Privacy-preserving machine learning

Background:

• Healthcare speech integration raises privacy concerns due to biometric data.
• Speaker anonymization protects personal information while preserving linguistic content.
• Anonymization of pathological speech remains underexplored despite critical privacy needs.

Purpose of the Study:

• To investigate the impact of speaker anonymization on pathological speech.
• To evaluate privacy enhancement, diagnostic utility, and demographic fairness.
• To explore deep-learning and signal processing anonymization methods.

Main Methods:

• Analysis of over 2700 speakers from multiple German institutions.
• Evaluation of deep-learning-based and signal processing-based anonymization techniques.
• Assessment of privacy (Equal Error Rate increase) and utility across various speech disorders.

Main Results:

• Substantial privacy improvements observed, with Equal Error Rate increases up to 1933%.
• Minimal impact on diagnostic utility for disorders like Dysarthria, Dysphonia, and Cleft Lip and Palate.
• Varying anonymization effects across disorders necessitate disorder-specific strategies; consistent fairness across demographics.

Conclusions:

• Speaker anonymization effectively enhances privacy in pathological speech.
• Customized, disorder-specific anonymization approaches are essential.
• Strategies must balance privacy with diagnostic utility and address potential inversion attacks.