Privacy-aware speaker trait and multimodal features relationship analysis in job interviews

Area of Science:

• Speech processing
• Biometrics
• Privacy-preserving machine learning

Background:

• Growing use of speech data for emotion detection and health profiling increases privacy risks.
• Voice recordings can reveal sensitive speaker traits, necessitating robust anonymization techniques.
• Maintaining speech utility for downstream tasks is crucial alongside identity protection.

Purpose of the Study:

• Investigate voice anonymization methods for protecting speaker identity.
• Assess the impact of anonymization on speech characteristics for trait inference, particularly in job interviews.
• Evaluate the trade-off between privacy and utility for different anonymization techniques.

Main Methods:

• Employed signal processing-based methods, including a phase vocoder.
• Utilized a neural audio codec-based method for voice anonymization.
• Conducted experiments to analyze acoustic parameter alterations and trait inference performance.

Main Results:

• Signal processing-based anonymization maintained speech suitability for trait assessment, comparable to original speech.
• The phase vocoder method provided modest privacy gains with acceptable utility trade-offs.
• The neural audio codec method altered critical prosodic features, slightly reducing trait estimation performance but offering greater privacy.

Conclusions:

• Signal processing methods offer a viable approach for voice anonymization in sensitive applications like job interviews.
• Carefully configured neural audio codecs can enhance privacy while preserving utility for speech recognition and quality assessment.
• Voice anonymization techniques must be carefully chosen based on the specific application's privacy and utility requirements.