ECAPA-Based Speaker Verification of Virtual Assistants: A Transfer Learning Approach

Speaker Verification technology is crucial in identifying individuals through their unique voice characteristics. However, the increasing use of speech assistants has posed new challenges for this technology. To address this, transfer learning is employed using the ECAPA-TDNN model trained on human...

Full description

Saved in:
Bibliographic Details
Published in2023 14th International Conference on Computing Communication and Networking Technologies (ICCCNT) pp. 1 - 7
Main Authors Rani, Drishti, Jain, Bhawna, Verma, Harshita, Varshney, Sona
Format Conference Proceeding
LanguageEnglish
Published IEEE 06.07.2023
Subjects
Computational modeling
Human voice
Internet
Speaker Recognition
Speech enhancement
SpeechBrain
Transfer learning
Virtual assistants
Voice Assistants
Online AccessGet full text

Cover

More Information
Summary:Speaker Verification technology is crucial in identifying individuals through their unique voice characteristics. However, the increasing use of speech assistants has posed new challenges for this technology. To address this, transfer learning is employed using the ECAPA-TDNN model trained on human voices from the VoxCeleb2 dataset. This study compares intra-voice assistant variations and proceeds to inter-voice assistant comparisons. In intra-pair comparisons, text-independent samples achieved accuracies of 83.33% (iOS versions) and 66.67% (Alexa versions), while text-dependent samples achieved 50% accuracy for both versions. In inter-pair comparisons (Alexa, Siri, Google Assistant, Cortana), accuracies of 100% (text-independent) and 80% (text-dependent) were observed. These findings showcase the effectiveness of transfer learning and the ECAPA-TDNN model for secure speaker verification in different speech assistant versions. The study provides valuable insights for enhancing speaker verification in the context of speech assistants.
ISSN:2473-7674
DOI:10.1109/ICCCNT56998.2023.10307339