Performance and Quality Evaluation of a McAdams speaker anonymisation for spontaneous German speech (vor Ort)
* Presenting author
Abstract:
A rapid increase in the use of commercial voice assistants has been observed during recent years due to the convenience of their usage across the age spectrum. Their advantage is an outstanding speech understanding also in difficult acoustic environments, their great popularity among the population, and their ease of use for information retrieval, which is enabled by cloud-computing AI-modelling. Typically, the use of voice assistants is limited to private users as the public use of voice assistants poses a threat to users’ identification especially as the user’s voice itself is seen as biometric information. An anonymization technique, such as the McAdams algorithm, can be applied to protect the biometric information of the user, by changing the voice impressions of the users. This approach intends to alter the speech features that define voice identity and allows for recordings of spontaneous voice data from interactions with voice assistants in conformity with the GDPR. Such an approach raises the question of the intelligibility and recognizability of anonymized speech. This paper aims at investigating the intelligibility and recognizability of anonymized speech across age and, gender on spontaneous German speech.