ADAPTIVE LOAD DISTRIBUTION IN VOICE IDENTIFICATION SYSTEMS

Abstract

This paper examines an approach to improving the efficiency of voice identification systems by distributing the computational load between the central processing unit (CPU) and the graphics processing unit (GPU). Existing implementations of such systems are based on fixed distribution schemes, within which individual processing stages are a priori associated with either the CPU or the GPU. Such a static organisation of computations does not take into account the dynamic variability of speech signal parameters, the current state of computing resources, and significant differences in the computational complexity of individual operations. Under such conditions, the system lacks mechanisms for adaptive load redistribution, resulting in asymmetric use of the CPU and GPU, and consequently, a decrease in overall performance. The proposed approach involves an adaptive algorithm that implements load distribution control based on the analysis of a set of speech recording characteristics, including frame length, overlap degree, energy content, and spectral saturation of the signal. The use of these parameters enables a quantitative assessment of the computational complexity of the current segment of the processed signal and the dynamic determination of the ratio of computational operations performed on the CPU and GPU. This ensures coordinated interaction between processors, minimises downtime, and increases the overall performance of the system. The comparative analysis revealed that the use of an adaptive algorithm significantly reduces the average processing time of speech fragments compared to approaches that rely solely on data volume. The proposed adaptive control and load distribution block increases the overall performance of voice identification systems, especially when processing large and structurally complex sets of signals, and can be integrated into modern multiprocessor architectures.