Excited to share our recent work on investigating the impact of voice anonymization on speech-based COVID-19 diagnostics!. Link: (1/).

RT @laurieallee: "Lasting damage to the heart and brain is an aspect of long COVID that should receive much more attention than it has so f….

RT @MuSAELab: Check out our latest paper "COVID-19 Detection via Fusion of Modulation Spectrum and Linear Prediction Speech Features" at IE….

Interested in our research in speech and cybersecurity? Consider following this account or our lab account @MuSAELab.

(7/7) .Q4: Can we protect privacy while maintaining high diagnostics accuracy?.ANSWER: Existing anonymization methods substitute the entire speaker embedding, making it difficult to do so. However, we found including anonymized speech samples in training data improves accuracy.

(6/7) .Q3: Why does diagnostics accuracy degrade?.ANSWER: The shifts of different speech features may give a clue. Less shift is seen with logmelspec ones, which are mostly correlated with speech content; more with the paralinguistic features (opensMILE/Modulation spectrum).

(5/7).Q2: Will this affect diagnostics accuracy?.ANSWER: Yes. The most severe degradation is observed when training and test data are anonymized using different methods (B&C). However, when the exact same methods are used, accuracy is much less affected (D).

(4/7) .Q1: Can anonymization tools effectively obfuscate user's identity? .ANWEWR: We experimented two types of anonymization methods. GAN-based approach is shown to be more effective than signal processing based method.

(3/7) Listen to anonymized audios here: .

(2/7) We argue that voice anonymization could potentially be a useful tool, to remove user's identity before the recording is upload to cloud.

(1/7) In existing speech diagnostics systems, voice recordings usually need to be uploaded to cloud where the large DNN models are deployed at. However, such process may lead to data privacy issues:

RT @MuSAELab: @BrennanSpiegel The Sensiks pod is very cool. We’ve quantified the impact of adding the different sensory stimuli on e.g., re….

RT @MuSAELab: Thank you for the shout out @LHonnorat ! It was a great event! @inrsciences.

ICASSP back to in-person :) and Rhodes island :)