A robust sequential test for text-independent speaker verification

Area of Science:

• Speech processing
• Machine learning
• Signal detection theory

Background:

• Speaker verification systems suffer performance degradation due to variations in training data, channel characteristics, and background noise.
• Traditional methods rely on empirical threshold setting, which may not be optimal under varying conditions.
• Robustness is crucial for reliable speaker verification in real-world scenarios.

Purpose of the Study:

• To develop a robust speaker verification algorithm that addresses performance degradation caused by data variability and noise.
• To introduce a minimax criterion for minimizing worst-case performance in speaker verification.
• To enhance decision confidence through sequential data processing.

Main Methods:

• Implementation of a sequential hypothesis testing framework for speaker verification.
• Application of a minimax criterion to optimize the detector's performance against a class of distributions.
• Utilizing sequential data acquisition to improve decision confidence when initial results are inconclusive.

Main Results:

• The developed sequential detector demonstrates robustness against common performance detractors in speaker verification.
• Performance was evaluated on a purpose-built, realistic database.
• The algorithm achieved performance comparable to or exceeding existing heuristic detection methods.

Conclusions:

• Sequential hypothesis testing offers a robust solution for speaker verification challenges.
• The minimax criterion enhances algorithm reliability under uncertain conditions.
• The proposed method provides a promising advancement for practical speaker verification systems.