Utility-Preserved Speech Anonymization via Generative Editing and Flow-Matching
September 4, 2026
A two-stage framework protects linguistic PII through generative speech editing and preserves acoustic identity using F3-VA, a flow-matching-based anonymization method. This approach maintains high utility for downstream ASR, TTS, and emotion recognition tasks compared to traditional methods.
HOW THIS AFFECTS YOU
●
builderYou can implement more effective privacy layers for voice-based products without destroying model utility.
●
policyThis offers a technical path toward complying with strict speech privacy regulations.