Research notes
Writings
Notes on the methods, experiments, and results from my speech AI security research.
-
2026
NaVo: Natural Voice Protection against Voice Cloning Attacks via Generative Universal Adversarial Audio
Generating rain, babble, and other background sounds that protect voices without optimizing a new perturbation for each recording. Published at Interspeech 2026.
Read article -
2025 – 2026
RoCo: Robust Code for Fast and Effective Proactive Defense against Voice Cloning Attack
Following RoVo, we moved voice protection into discrete codec tokens to reduce generation time while retaining resistance to speech enhancement.
Read article -
2024 – 2025
RoVo: Robust Voice Protection Against Voice Cloning Attacks via Embedding-Level Adversarial Perturbations
Voice protection in a neural audio codec's latent space, tested against signal transforms, speech enhancement, reconstruction, and unseen cloning models.
Read article -
2023 – 2024
Voice Synthesis Detection Using Language Model-Based Speech Feature Extraction
Using Encodec tokens and a pretrained BERT model to detect synthetic speech, with experiments on ASVspoof 2019 and 2021.
Read article -
2022 – 2023
Comparison of Korean Speech De-identification Performance of Speech De-identification Model and Broadcast Voice Modulation
A comparison of pitch shifting, McAdams, resampling, and VTLN on Korean call-center recordings, with listening tests and speaker-verification experiments.
Read article
