بازشناسي گوينده در گفتار تلفني زبان فارسي با استفاده از الگوريتم‌هاي يادگيري عميق

Speaker recognition , telephone speech , laboratory speech , deep learning algorithms an‎d score normalization

Speaker Recognition in Persian Telephone Speech Using Deep Learning Algorithms

زبان شناسي

Abstract Speaker recognition in the context of telephone speech faces the challenge of "amplitude mismatch" compared to laboratory speech due to ban‎dwidth limitations, real network noise, an‎d channel distortions. This leads to a noticeable decline in the performance of conventional systems. The aim of this research is to implement an‎d eva‎luate a speaker recognition system for the Persian language an‎d to study its behavior in the face of telephone speech conditions an‎d domain mismatch. In this regard, a deep learning-based approach an‎d speaker embedding vector extraction were employed, utilizing the ECAPA-TDNN an‎d ResNet-34 architectures. The issue was eva‎luated both in terms of speaker identification (multi-class classification) an‎d speaker verification (similarity-based decision-making). To reduce the effect of domain mismatch, the models were trained using clean laboratory data an‎d then eva‎luated on real telephone speech data. In the detection section, the cosine criteria, ROC curves, an‎d DET curves yielded the AUC value an‎d equal error rate. The effect of a post-processing step of normalizing the scores on the eva‎luation was also confirmed an‎d reviewed. The results showed that both models experienced a decline in performance when faced with telephone data, indicating the significant role of domain mismatch. However, the ECAPA-TDNN architecture, utilizing a channel attention mechanism an‎d multi-scale aggregation, demonstrated more stable performance than ResNet-34, achieving higher recognition accuracy an‎d a lower equal error rate. Additionally, speaker verification eva‎luation showed that applying Z-norm as an eva‎luation supplement can regularize the score distribution an‎d improve equal error rate an‎d AUC metrics in both models. Overall, the findings emphasize that for telephone applications, choosing the appropriate architecture an‎d using score normalization methods are effective strategies for reducing the adverse effects of the channel an‎d achieving a more robust system for Persian speaker recognition. Keywords: Speaker recognition, telephone speech, laboratory speech, deep learning algorithms an‎d score normalization

5

161857