Challenge of speech anonymization: cannot develop a model which both preserves speech features well but also effectively anonymizes the speech.
Methodology
- Separate content and speech encoders
- Results in highly concentrated + effective speech content, but with very widespread voiceprint