A research framework called NouveauVoice can replace your real voice with a convincing fake one that automated systems struggle to trace back to you.
The system, detailed in a new arXiv preprint, builds on a Hierarchical Deep Variational Autoencoder to generate what the authors call pseudo-speaker embeddings - synthetic voice profiles that sound like real people but don't match anyone in a verification database. It slots in as a plug-in on top of two existing speech architectures, FACodec and CosyVoice2, rather than replacing them outright. Tested against an automatic speaker verification attacker, it achieved an Equal Error Rate above 38% - meaning the verifier was nearly guessing at random when trying to identify the original speaker. The researchers also found the anonymized audio retained intelligibility and emotional expressiveness, two things earlier anonymization tools tended to crush.
This matters because voice cloning and conversion tools have made it trivially easy to impersonate someone from a short audio clip, and most speaker anonymization approaches have lagged behind by working in feature spaces too narrow to generate truly distinct fake identities. NouveauVoice directly targets that gap by using probabilistic sampling to push synthesized voices further apart from each other and from the source.
The catch: a 38% equal error rate is a meaningful result in a lab benchmark, but real-world voice verification systems are a moving target - and the same generative methods powering NouveauVoice are also what makes voice spoofing so potent in the first place.