Anonymization, Not Elimination: Utility-Preserved Speech Anonymization
A novel two-stage framework replaces PII and anonymizes voices while preserving data utility for ASR and TTS models.
A team of researchers has introduced a novel framework for speech anonymization that aims to solve a critical tension in AI development: protecting user privacy without destroying the value of the data for training models. The paper, 'Anonymization, Not Elimination: Utility-Preserved Speech Anonymization,' proposes a two-stage approach. First, a generative speech editing model seamlessly replaces personally identifiable information (PII) within the audio to protect linguistic content. Second, a new flow-matching-based framework called F3-VA anonymizes the speaker's voice, generating diverse and distinct synthetic speakers to protect acoustic identity.
Current anonymization methods often severely degrade data quality, disrupting acoustic continuity or reducing vocal diversity, which harms the performance of models trained on that data, such as Automatic Speech Recognition (ASR) or Text-to-Speech (TTS) systems. The researchers also critique existing evaluation practices, which typically just test anonymized speech on pre-trained models. Instead, they propose a more rigorous protocol that assesses utility by training ASR, TTS, and Speech Emotion Recognition (SER) models from scratch on the anonymized data.
Experimental results show the framework achieves stronger privacy protection than established baselines from the VoicePrivacy Challenge while causing minimal utility degradation. This means companies and researchers could use this technique to create large, privacy-compliant speech datasets that remain highly useful for building and improving core speech AI technologies, potentially accelerating development in a privacy-conscious era.
- Uses a two-stage framework: generative editing replaces PII for content privacy, and the F3-VA model anonymizes speaker voice.
- Outperforms VoicePrivacy Challenge baselines, providing stronger privacy with minimal loss of data utility for model training.
- Proposes a new evaluation protocol that trains models like ASR and TTS from scratch, giving a more realistic measure of utility.
Why It Matters
Enables the creation of large, privacy-safe speech datasets that don't compromise the performance of AI models trained on them, balancing innovation with ethical data use.