JMIR Publications

ABSTRACT

Background:

Large, curated datasets are required to leverage speech-based tools in healthcare. These are costly to produce, resulting in increased interest in data sharing. As speech can potentially identify speakers (i.e., voiceprints), sharing recordings raises privacy concerns. This is especially relevant when working with patient data protected under HIPAA.

Objective:

We aimed to determine the re-identification risk for speech recordings, without reference to demographics or metadata, in clinical datasets, considering both the size of the 'search space', i.e., the number of comparisons that must be considered when re-identifying, and the nature of the speech recording, i.e., the type of speech task.

Methods:

Using a state-of-the art speaker identification model, we model an adversarial attack scenario where-in an adversary uses a large dataset of identified speech (hereafter the 'known' set) to re-identify as many unknown speakers in a shared dataset (hereafter the 'unknown' set) as possible. We first considered the effect of search space size by attempting re-identification with various sizes of known and unknown sets using VoxCeleb, a dataset with recordings of natural, connected speech from over 7000 healthy speakers. We then repeated these tests with different types of recordings in each set to examine whether the nature of a speech recording influences re-identification risk. For these tests, we used our clinical dataset composed of recordings of elicited speech tasks from 941 speakers.

Results:

We found that the risk is inversely related to the number of comparisons an adversary must consider, i.e., the ‘search space’, with a positive linear correlation between the number of false acceptances and number of comparisons (r = 0.69, P < .001). In effect, risk is high for a small search space but drops as the search space grows (precision > 0.85 for < 1∗10^6 comparisons, precision < 0.5 for > 3∗10^6 comparisons. We also found that the nature of a speech recording influences re-identification risk, with non-connected speech (e.g., vowel prolongation, precision = 0) being harder to identify than connected speech (e.g., sentence repetition, precision = 0.665).

Conclusions:

Our findings suggest that speaker identification models can be used to re-identify participants in specific circumstances, but in practice, the re-identification risk appears small. The variation in risk due to search space size and type of speech task additionally provides actionable recommendations to further increase participant privacy and considerations for policy around public release of speech recording.