STUDY OF NOISE REDUCTION METHODS IN THE SOUND SEQUENCE WHEN SOLVING THE SPEECH-TO-TEXT PROBLEM

Abstract

The subject of this research is noise reduction methods in the sound sequence as a part of the proposed speech-to-text (STT) module for converting a verbal lecture or a lesson into a written text form on digital educational platforms. The goal is to investigate the influence of noise reduction methods on the operation of the acoustic signal recognition system. 3 methods of noise reduction were considered for integration in the proposed acoustic artifact recognition system and for the researching: spectral subtraction method; fast Fourier transform; Wiener filter with software modeling of every method. The obtained results: after testing the system with integrated noise reduction modules in it, based on the fast Fourier transform, Wiener filter and spectral subtraction method, it was concluded that the module using the Wiener filter improves the identification results by 25%, which is the highest result. However, performance testing has shown that fast Fourier transform is the fastest method. The practical significance of the work is – the identifying acoustic events system was developed, different noise reduction methods were integrated and researched into the module for converting a verbal lecture or a lesson into a written text form in a proposed system with the aim of increasing of speed and accuracy.