Content constraints improve reversible speaker anonymization against unseen verifiers
Original title: The Content Leash: Mitigating Degenerate Solutions in Invertible Speaker Anonymization
On 38 unseen speakers, attacker equal error rate rose from 2.8% to 26.3%, with 11.8% word error rate.
The paper adds content-preserving training constraints and a stronger speaker-verification representation to reversible voice anonymization. The goal is to hide identity without relying on speech degradation and to remain effective against verifiers not used in the original setup. Authorized users can still restore the original voice. Reported results cover unseen speakers and a held-out verifier while keeping intelligible speech and high restoration similarity.
- Equal error rate reaches 27.9% against a held-out verifier.
- Key-based restoration retains 94.9% speaker similarity.
- The measured attacker error-rate floor before the method was 2.8%.
- Word error rate after the approach is 11.8%.
From the original article
1 min read at arxiv.orgVoice anonymization aims to protect a speaker's identity while preserving the linguistic content of speech, enabling voice data to be shared and processed with reduced privacy risk. This work focuses on reversible voice anonymization, where the original voice can be recovered by an authorized user rather than being permanently discarded.
By Nirjon Mobarak Talukder · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.