ABGESAGT - Investigation of acoustic features in selection of optimal target speakers for speaker anonymisation (en)

* Presenting author
Day / Time: 09.03.2023, 11:00-11:40
Typ: Poster
Information: Die Poster werden durchgängig von Dienstag bis Donnerstag im Foyer auf Ebene Z ausgestellt (3.OG vor dem Plenarsaal Z), sortiert nach thematischen Zusammenhängen in der hier angegebenen Posterinsel (siehe Sitzungstitel). Das Posterforum zur angegebenen Zeit bietet die Möglichkeit, mit den Autor*innen in Diskussion zu kommen.

Speaker anonymization aims at hiding a speaker's identity while preserving prosodic and linguistic content. In this work, a voice conversion based speaker anonymization is used, where a source speaker's voice is transformed to that of the target speaker. We observe that the choice of a suitable target speaker is crucial for successful anonymization. If the chosen target speaker is already similar to the source speaker, the anonymization fails with no fault of the anonymization process, as favorably the most dissimilar target speaker for a source speaker should be chosen. However, for every new source speaker, the model could fail to find one. Therefore, we examine the prosody-related acoustic features, such as F0, power envelope, etc., instead of only relying on the speaker dissimilarity, known as speaker embeddings, which might not be available during the anonymization process. While speaker embeddings are model dependent, acoustic features are independent of the anonymization model and ergo, aide in finding optimal target speaker.



strpos(): Passing null to parameter #1 ($haystack) of type string is deprecated

/kunden/431424_13355/webseiten/subdomains/, line 35
32.              */
33.             if (headers_sent()) {
34.                 $encoding = false;
35.             } elseif (strpos($HTTP_ACCEPT_ENCODING, 'x-gzip') !== false) {
36.                 $encoding = 'x-gzip';
37.             } elseif (strpos($HTTP_ACCEPT_ENCODING, 'gzip') !== false) {
38.                 $encoding = 'gzip';