Logo des Repositoriums

Auflistung nach Schlagwort "Speech and speaker recognition"

1 - 2 von 2
  • Konferenzbeitrag
    Assessing the Human Ability to Recognize Synthetic Speech in Ordinary Conversation
    (BIOSIG 2023, 2023) Daniel Prudký, Anton Firc
    This work assesses the human ability to recognize synthetic speech (deepfake). This paper describes an experiment in which we communicated with respondents using voice messages. We presented the respondents with a cover story about testing the user-friendliness of voice messages while secretly sending them a pre-prepared deepfake recording during the conversation. We examined their reactions, knowledge of deepfakes, or how many could correctly identify which message was deepfake. The results show that none of the respondents reacted in any way to the fraudulent deepfake message, and only one retrospectively admitted to noticing something specific. On the other hand, a voicemail message that contained a deepfake was correctly identified by 83.9% of respondents after revealing the nature of the experiment. Thus, the results show that although the deepfake recording was clearly identifiable among others, no one reacted to it. In summary, we show that the human ability to recognize voice deepfakes is not at a level we can trust. It is very difficult for people to distinguish between real and fake voices, especially if they do not expect them.
  • Konferenzbeitrag
    Fairness and Privacy in Voice Biometrics: A Study of Gender Influences Using wav2vec 2.0
    (BIOSIG 2023, 2023) Oubaida Chouchane, Michele Panariello
    This study investigates the impact of gender information on utility, privacy, and fairness in voice biometric systems, guided by the General Data Protection Regulation (GDPR) mandates, which underscore the need for minimizing the processing and storage of private and sensitive data, and ensuring fairness in automated decision-making systems. We adopt an approach that involves the fine-tuning of the wav2vec 2.0 model for speaker verification tasks, evaluating potential gender-related privacy vulnerabilities in the process. An adversarial technique is implemented during the fine-tuning process to obscure gender information within the speaker embeddings, thus bolstering privacy. Results from VoxCeleb datasets indicate our adversarial model increases privacy against uninformed attacks (AUC of 46.80\%), yet slightly diminishes speaker verification performance (EER of 3.89\%) compared to the non-adversarial model (EER of 2.37\%). The model's efficacy reduces against informed attacks (AUC of 96.27\%). Preliminary analysis of system performance is conducted to identify potential gender bias, thus highlighting the need for continued research to understand and enhance fairness, and the delicate interplay between utility, privacy, and fairness in voice biometric systems.