The VoicePrivacy 2022 Challenge: Progress and Perspectives in Voice Anonymisation (2407.11516v1)

Abstract: The VoicePrivacy Challenge promotes the development of voice anonymisation solutions for speech technology. In this paper we present a systematic overview and analysis of the second edition held in 2022. We describe the voice anonymisation task and datasets used for system development and evaluation, present the different attack models used for evaluation, and the associated objective and subjective metrics. We describe three anonymisation baselines, provide a summary description of the anonymisation systems developed by challenge participants, and report objective and subjective evaluation results for all. In addition, we describe post-evaluation analyses and a summary of related work reported in the open literature. Results show that solutions based on voice conversion better preserve utility, that an alternative which combines automatic speech recognition with synthesis achieves greater privacy, and that a privacy-utility trade-off remains inherent to current anonymisation solutions. Finally, we present our ideas and priorities for future VoicePrivacy Challenge editions.

An Examination of the VoicePrivacy 2022 Challenge: Advancing the Field of Voice Anonymization

The paper "The VoicePrivacy 2022 Challenge: Progress and Perspectives in Voice Anonymization" presents a comprehensive overview of the second edition of the VoicePrivacy Challenge, targeting the development of advanced voice anonymization techniques. The aim of this paper is to highlight the methodologies, results, and future directions gleaned from the 2022 challenge, which has crucial implications for privacy preservation in speech technologies.

The challenge’s scenario involves two main actors: a user who desires to anonymize their voice recordings before sharing and an adversary aiming to de-anonymize these recordings. This scenario emphasizes the importance of robust voice anonymization techniques that can effectively thwart identification attempts while maintaining the utility of the anonymized recordings.

Overview and Protocol

The challenge was framed around several key aspects:

• Anonymization Task: Participants were required to develop systems that output anonymized speech waveforms, mask speaker identities, and preserve linguistic and paralinguistic attributes. Consistency in anonymizing speech attributed to the same speaker while distinguishing different speakers was also a requirement.
• Attack Models: Evaluation was based on several attack models where adversaries could leverage anonymization-aware techniques to de-anonymize recordings.
• Datasets: Publicly available datasets including VoxCeleb, LibriSpeech, and VCTK were used for system development and evaluation, ensuring consistency and comparability among submissions.

Metrics for Evaluation

A dual approach using objective and subjective metrics was adopted:

1. Privacy Metrics: Evaluated through ASV systems with metrics such as Equal Error Rate (EER).
2. Utility Metrics: Assessed via word error rates (WER) using ASR systems.

Secondary metrics such as pitch correlation (ρF0) and gain in voice distinctiveness (GVD) were introduced to gauge prosody preservation and consistency in pseudo-voices.

Baseline and Submitted Systems

Three baseline systems (B1.a, B1.b, and B2) were provided, each leveraging different techniques:

• B1.a and B1.b employed x-vectors and neural waveform models, with variations in bottleneck features and synthesis methods.
• B2 used a signal processing approach focused on LPC and McAdams coefficients for formant modification.

Participants submitted varied approaches mainly revolving around these baselines with enhancements:

• T04: Utilized a TTS-based approach providing strong privacy protection but slightly lower pitch correlation.
• T11: Implemented advanced feature extraction and speaker embedding techniques, balancing privacy and utility well.
• T18: Focused on enhancing anonymization functions using adversarial noise and alternative speaker embeddings.
• T40: Improved pitch alignment with the use of F0 regression models.
• T32: Leveraged signal processing technologies reminiscent of the B2 approach.

Results and Analysis

From the objective metrics, T11 and T04 achieved notable improvements in EER and WER, indicating robust privacy and utility performance, respectively. T11 systems demonstrated an ability to finely tune pseudo-speaker similarity, directly affecting voice distinctiveness (GVD).

Subjective evaluations of naturalness and intelligibility highlighted that despite the trade-offs, advanced systems like T11 and T04 were able to maintain competitive naturalness and intelligibility scores compared to baselines. The verifiability scores showed high overlap between target and non-target trials, suggesting effective anonymization against human listeners.

Implications and Future Directions

The implications of these advancements are significant for both practical and theoretical aspects of voice anonymization:

• Practical: Improved anonymization techniques can substantially enhance user privacy in real-world applications, from social media interactions to sensitive voice-based services.
• Theoretical: The integration of detailed evaluation metrics like pitch correlation and voice distinctiveness helps in refining anonymization techniques and understanding their limitations and strengths.

Future developments anticipated from these findings include:

• Enhancing the integration of TTS-based anonymization approaches while maintaining prosody.
• Refining privacy metrics to better capture real-world attack scenarios and adversarial capabilities.
• Developing lightweight, real-time anonymization solutions that can operate efficiently in constrained environments.

The paper highlights that while significant progress has been made, the field continues to evolve. Future editions of the VoicePrivacy Challenge will aim to incorporate more holistic evaluation frameworks and usability improvements to foster further advancements in voice anonymization technology.