کنفرانس بین المللی فناوری اطلاعات و دانش

English

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Listening with Precision: ASR-Guided Method and Fusion Strategy for Text-Dependent Speaker Verification

نویسندگان :

Mohammad Reza Molavi¹ Reza Khodadadi² Hossein Zeinali³

1- دانشگاه صنعتی امیرکبیر (پلی‌تکنیک تهران) 2- دانشگاه صنعتی شریف 3- دانشگاه صنعتی امیرکبیر (پلی‌تکنیک تهران)

کلمات کلیدی :

speaker verification،text-dependent, ASR،speaker embedding fusion

چکیده :

This paper proposes a text-dependent speaker verification approach (TD-SV) that improves accuracy and robustness by leveraging automatic speech recognition (ASR) to guide the verification process and final fusion score. Our system inte- grates a Fast-Conformer-based ASR module to validate speech content, effectively filtering out target-wrong and impostor- wrong trials. We propose a feature fusion method for speaker verification that combines speaker embeddings from Wav2Vec- BERT and ReDimNet, leveraging self-supervised and task- specific representations for superior performance. This fusion significantly improves verification accuracy compared to indi- vidual embeddings. Our approach achieves a competitive nor- malized minDCF of 0.045 on the Iranian division of the TD- SV 2024 Challenge test set, demonstrating an optimal balance between performance and computational efficiency. Our best submission secured the second rank in the challenge.