ACOUSTIC INVARIANT APPROACH TO SPEECH SOUND ANALYSIS FOR BRAND NEW SPEECH RECOGNITION SYSTEMS (UKRAINIAN AND ENGLISH)

Maksym O Vakulenko

doi:10.47839/ijc.6.3.455

Abstract

On the basis of acoustic invariant speech analysis (AISA), the permanent spectral characteristics of the Ukrainian vowels are obtained for various ways of pronunciation including ordinary speech, whisper and changing tone. It is shown that the lowest phonemic frequencies due to vocal fold oscillations or to Helmholtz resonance are not associated with persistent sound features. It is conjectured that the only phonemic invariant is the ratio between formant frequencies, not their absolute values. This analysis is complemented by the computer sound synthesis. We show also that the acoustic invariants of the Ukrainian sound [i] are close to that of English [I]. The results obtained may be useful for specialists in the field of experimental phonetics and speech modelling.

Highlights

Investigation of human speech sounds is a key task in experimental phonetics and related branches, as well as in various fields of applied science and engineering
We show that the acoustic invariants of the Ukrainian sound [i] are close to those of English [I]
The results obtained may be useful for specialists in the field of experimental phonetics and speech modelling

Summary

INTRODUCTION

Investigation of human speech sounds is a key task in experimental phonetics and related branches, as well as in various fields of applied science and engineering. Being a much more explored field than its acoustic counterpart, the articulatory approach is widely used in speech recognition systems that are based, in most cases, on the so-called generative model [29] Within this model, any incoming sound signal is compared to a set of machine-generated reference waveforms produced on the basis of knowledge about human articulation. Any incoming sound signal is compared to a set of machine-generated reference waveforms produced on the basis of knowledge about human articulation Not surprisingly, this approach makes it impossible to separate speaker-independent acoustic characteristics responsible for formation of a given speech sound from ones reflecting the talker’s individuality. The acoustic invariants found for the vowels analyzed, should be taken into account in modern talker-independent speech recognition software

Acoustic invariant speech analysis: background and methodology

Speech sound analysis and synthesis

Summary