Preprint
Noise Masking Attacks and Defenses for Pretrained Speech Models
Matthew Jagielski et al., arXiv, 2024
Read the originalRead on October 17, 2026
- What it is
- A 2024 preprint from Google researchers showing that speech models trained on sensitive audio can give back the memorized transcript of a training recording when the same recording is sent with its words partly masked by noise.
- How much weight it carries
- A preprint, not yet peer-reviewed in the version read. The authors come from a major speech research group and the method is described in detail. The attacks use models built for the study, and the success rate depends on how long a model is trained.
Facts taken from this source
- A 2024 study found that when a speech model has seen a recording during training, a noisy version of that recording can make the model transcribe its memorized sensitive transcript.
“They show that when a record has been seen at training time, the model will transcribe the noisy record with its memorized sensitive transcript.” (Abstract)
- The authors of a 2024 study state that speech models are often trained on sensitive data to improve their performance, which can lead to privacy leakage.
“Speech models are often trained on sensitive data in order to improve model performance, leading to potential privacy leakage.” (Abstract)