Self-Supervised Open-Set Speaker Recognition with Laguerre-Voronoi Descriptors.

Area of Science:

• Behavioral Biometrics
• Machine Learning
• Signal Processing

Background:

• Speaker recognition is a critical but challenging area within behavioral biometrics.
• Existing deep learning methods primarily focus on closed-set systems, leaving open-set recognition under-explored.
• Robust speaker verification in real-world scenarios remains a significant hurdle.

Purpose of the Study:

• To propose a novel self-supervised open-set speaker recognition framework.
• To enhance speaker verification accuracy and robustness by leveraging geometric properties of speaker distribution.
• To address limitations in current state-of-the-art open-set speaker recognition systems.

Main Methods:

• Development of a deep neural network (DNN) incorporating a wider temporal speech feature perspective.
• Implementation of Laguerre-Voronoi diagram-based speech feature extraction.
• Training the DNN using a specialized clustering criterion that requires only positive pairs.

Main Results:

• The proposed self-supervised system demonstrated superior performance compared to existing state-of-the-art methods.
• Significant improvements were observed in open-set speaker recognition accuracy.
• Enhanced cluster representation capabilities were validated through experimental results.

Conclusions:

• The proposed self-supervised framework offers a promising solution for accurate and robust open-set speaker recognition.
• Leveraging geometric properties of speaker distribution is effective for improving speaker verification.
• The method advances the field of behavioral biometrics and machine learning applications in speaker identification.