Key points are not available for this paper at this time.
One of the most popular areas of research in voice technology is speaker recognition. There has been a noticeable shift in recent years toward entrenching based end-to-end speaker recognition techniques. In real-world situations, these techniques allow speaker recognition without requiring retraining when faced with unfamiliar voices. Previous models trained on tiny, repetitive, hand-crafted datasets performed well in pattern corresponding but poorly in robustness due to architectural limitations and data scarcity. It has been demonstrated that training attention-based deep neural networks—which use large datasets with a variety of incorrectly classified examples—can greatly increase their robustness when used in a variety of classification tasks later on. Using insights from large-scale datasets such as VoxCeleb1 and VoxCeleb2, in addition to custom-crafted proprietary data designed to accommodate several modalities, we began exploring the potential of these flexible approaches to develop speaker identification models. More specifically, we looked at two well-known models to see how the architectural improvements might affect the model's capacity for effective generalization: Titanet 2 and ECAPA-TDNN 1 (highlighted Channels Pay attention, Transmission, and Aggregate). We also suggest utilizing Elastic search to create a scalable and quicker inference process.
Rathod et al. (Fri,) studied this question.