Comparison of Cubic and Fibonacci Activation Functions in Speech Signal Seperation using Independent Component Analysis

Authors

  • James P. Chibole
  • Heywood A. Ouma
  • Edward Ndungu

Keywords:

Blind Signal Processing (BSS), Independent Component Analysis (ICA), Natural Gradient Algorithm (NGA), Activation Functions, magnitude-squared coherence

Abstract

The cubic activation function has been used extensively in audio signal separation algorithms. The intention of this paper is to measure the quality of separation for this activation function and compare its performance to that of the Fibonacci. The setup entails the use of the Natural Gradient Algorithm (NGA) to separate two mixed signals into their original components using the Independent Component Analysis (ICA). The NGA used is formulated using instantaneous Blind Signal Processing (BSP). The design uses a 2 x 2 Multiple Input Multiple Output (MIMO) system to accept the two speech signals, mix them and separate them to retain their original form or their filtered version. Two activation functions; the Cubic and Fibonacci are used interchangeably. The results are compared by analyzing the magnitude-squared coherence of the input and separated signals. The results show that Fibonacci is best suited to separate speech signals when applied in NGA than the cubic
activation function.

Author Biographies

James P. Chibole

Jomo Kenyatta University of Agriculture and Technology.

Heywood A. Ouma

University of Nairobi.

Edward Ndungu

Jomo Kenyatta University of Agriculture and Technology.

Downloads

Published

04-04-2022