UPDATING USER PROFILES METHODS IN MULTIVECTOR VOICE IDENTIFICATION SYSTEMS

Abstract

The article presents the results of a comparative analysis of approaches to updating and adapting vector user profiles in speech identification systems, along with a new method. The article focuses on maintaining the stability and accuracy of recognition, taking into account voice changes between sessions, due to physiological factors, variations in articulation and acoustic recording conditions. Methods such as incremental vector addition, partial and complete feature replacement, and multi-vector approaches with fixed or adaptive weighting are analysed.

The proposed method assigns a weight coefficients based on the similarity between new voice samples and existing profile vectors, thereby balancing adaptability with robustness to noise and atypical voice samples. Experiments were conducted to evaluate different mechanisms for updating user profiles and to assess their impact on the accuracy, stability, and robustness of identification.

Experimental results show that multi-vector approaches with distinct weighting schemes, including the proposed method, more effectively account for vocal changes between sessions and variable acoustic conditions. This process reduces the false-rejection rate compared to fixed-profile and single-vector updating approaches. The choice of update mechanism significantly influences the dynamics of profile adaptation, system stability in the presence of noise and physiological voice changes, and the balance between recognition accuracy and computational costs.

The results of this development have practical significance for the design and configuration of speech recognition systems that rely on user profiles for long-term performance. This paper offers recommendations for selecting an update approach based on operational conditions and the desired balance between recognition accuracy, robustness, and computational efficiency.