Adjustable Deterministic Pseudonymization of Speech

Obfuscating Voice Identity

As public speech resources become increasingly available, there is a growing interest to preserve the privacy of the speakers, while preserving the spoken message. Researchers from Idiap Research Institute and the Netherlands Cancer Institute are collaborating to develop a method for pseudonymization -- reversible anonymization -- of speech, that allows to obfuscate the speaker identity in untranscribed running speech. Briefly, the proposed pseudonymization approach modifies the spectro-temporal structure by simulating a different length and structure of the vocal tract system response and altering the pitch frequency and speaking rate. The resultant method is deterministic and partially reversible, and the changes are adjustable on a continuous scale.




The method is implemented in Praat and is publicly available.

Below is a demo:


Original 1

Pseudonymized 1


Original 2

Pseudonymized 2

This work was partially funded by the Hasler Foundation under the project Flexible Linguistically-guided Objective Speech assessment (FLOSS) and by Innosuisse under the project Conversation Member Match (CMM).