بازشناسي گوينده در گفتار بداهه فارسي با استفاده از يادگيري عميق

Speaker recognition under spontaneous speech conditions represents one of the most challenging problems in speech processing, as intra-speaker variability arising from fluctuations in tone, speech rate, emotional states, an‎d environmental conditions significantly hinders the extraction of stable speaker identity representations. This challenge is further compounded in the Persian language due to limited training resources an‎d dialectal diversity. The present study conducts a comparative eva‎luation of three families of deep learning models — SincNet, RawNet, an‎d Wav2Vec2 — for speaker recognition from Persian spontaneous speech. The speech corpus comprises recordings from 56 male Persian-speaking participants produced in a fully text-independent, spontaneous style under uncontrolled environmental conditions. To realistically simulate deployment scenarios, a chronological data split was employed in lieu of ran‎dom partitioning, an‎d all utterances were segmented into fixed two-second segments. All models were trained using Additive Angular Margin Softmax loss to maximize inter-class separation an‎d minimize intra-class variance. Results demonstrate that the Wav2Vec2-large-xlsr-53 model, augmented with a Speaker-Aware Attention Pooling mechanism, substantially outperforms all competing approaches, achieving 99.58% accuracy on the held-out test set. The SincNet variant ranks second at 92.52%, establishing itself as a lightweight an‎d interpretable alternative for resource-constrained deployment environments. RawNet2 an‎d RawNet3 attain 86.34% an‎d 83.32% accuracy, respectively. The findings confirm that multilingual self-supervised pre-training combined with dynamic attention-based pooling constitutes the most effective strategy for managing intra-speaker variability in Persian spontaneous speech. Furthermore, the study demonstrates that chronological data partitioning provides a more realistic performance estimate by preventing statistical overlap between training an‎d test distributions — a common source of inflated results in ran‎domly split eva‎luations. These results establish a rigorous benchmark for speaker recognition in low-resource Persian speech an‎d offer actionable insights for future research on robust speaker modeling in real-world acoustic conditions.

5

167551