Software system for speaker verification using parallel CPU and GPU computing

Abstract: This paper is devoted to speaker verification software using parallel CPU and GPU computing. This software is based on Gaussian mixture model and universal background model (GMM-UBM system). Developed software allows to train the universal background model (UBM), speaker models and test recorded speech samples in order to verify their belonging to the selected speaker model. Also, software provides speech feature selection module using greedy add-del and genetic algorithms. The experimental evaluation of the UBM training module was conducted using CPU, GPU and combined parallel calculations. Parallel CPU and GPU calculations results in 36,95% calculations time decrease compared to parallel CPU implementation, and 10% decrease compared to only GPU implementation.

Keywords: speaker recognition, speaker verification, gmm-ubm system, speech processing, software system, parallel computations, gpu, cuda, gaussian mixture model

For citation:
Rahmanenko I. A. Software system for speaker verification using parallel CPU and GPU computing. Doklady Tomskogo gosudarstvennogo universiteta sistem upravleniya i radioelektroniki, 2017, vol. 20, no. 1, pp. 70–74. DOI: 10.21293/1818-0442-2017-20-1-70-74