LOCUST - Longitudinal Corpus and Toolset for Speaker Verification
Résumé
In this paper, we set forth a new longitudinal corpus and a toolset in an effort to address the influence of voice-aging on speaker verification.
We have examined previous longitudinal research of age-related voice changes as well as its applicability to real world use cases. Our findings reveal that scientists have treated age-
related voice changes as a hindrance instead of leveraging it to the advantage of the identity validator. Additionally, we found a significant dearth of publicly available corpora related to
both the time span of and the number of participants in audio recordings. We also identified a significant bias toward the development of speaker recognition technologies applicable to
government surveillance systems compared to speaker verification systems used in civilian IT security systems.
To solve the aforementioned issues, we built an open project with the largest publicly available longitudinal speaker database, which includes 229 speakers with an average talking
time exceeding 15 hours spanning across an average of 21 years per speaker. We assembled, cleaned, and normalized
audio recordings and developed software tools for speech features extractions, all of which we are releasing to the publi