Adjustable deterministic pseudonymization of speech

While public speech resources become increasingly available, there is a growing interest to preserve the privacy of speakers, through methods that anonymize speaker information from while preserving spoken linguistic content. In this paper, method for pseudonymization (reversible anonymization) presented, allows obfuscate identity in untranscribed running speech. The approach manipulates spectro-temporal structure simulate different length and vocal tract by modifying formant locations, as well altering pitch speaking rate. deterministic partially reversible, changes are adjustable on continuous scale. has been evaluated terms (i) ABX listening experiments, (ii) automatic verification recognition. experimental results indicate identifiability among forced choice pairs reduced over 90% less than 70% pseudonymization, de-pseudonymization was effective. An evaluation VoicePrivacy 2020 challenge data showed proposed performs better signal processing based baseline uses McAdams coefficient slightly worse neural source filtering method. Further analysis approach: comparable phone posterior feature objective intelligibility measure, preserves tracks method, (iii) paralinguistic aspects such dysarthria several speakers.