Distribution of Operations in Heterogeneous Computing Systems for Processing Speech Signals

Abstract

The exponential growth of data is forcing the search for new approaches to computing power. The diversity of data is increasing, and with it is the need for advanced techniques such as artificial intelligence (AI), machine/deep learning to help transform that data into information. Speech signal processing in particular is one of them. As a solution, generic computing is being replaced by heterogeneous computing. This article describes the technologies of parallel processing and distributed operations of spectral transformation of speech signals using central processing unit (CPU) and graphics processing unit (GPU). The one problem of parallel processing of spectral transformation of speech signals is imbalance among the operations between CPU and GPU which leads to performance degradation. A serious problem with spectral transform is the selection of the appropriate frame size of the speech signal for parallel processing on the CPU or GPU. The article also proposes a fast algorithm for spectral transformation of speech signals using OpenMP and CUDA technologies, and results of the influence of the number of threads and the frame size of the speech signal on the acceleration is also shown.