Time-frequency representation of speech signal extraction of acoustic parameters

摘要

The non - stationary nature of speech signal requires, for its representation, the use of a time frequency type description. Indeed, the evolution in the time domain of its spectrum is a necessary condition for the acoustic signal to carry a linguistic code. The spectral analysis based on the Fourier transform is not adequate because it does not provide sufficient information about the time evolution of the spectrum. The purpose of this paper is to put in evidence the acoustic parameters evolution of a speech signal (formants of some vowels of the French language) in time and in frequency based on more specific tools, adapted to non-stationary signal analysis. Such tools, grouped under the name of time - frequency representation, try to represent more accurately the evolution of the spectral contents of the signal as a function of time. For implementation, in random cases, a construction under the constraint of time - frequency representation based on the extension of the stationary notion of the spectral density is possible. For such a construction, two approaches are available, either preserving the frequency visual concept, or keeping the double orthogonal decomposition. We will start with linear representations and then move on to the quadratic energy representations based on the Cohen class which regroup the invariant distributions by time and frequency translations. The role of Wigner - Ville distribution and the attenuation of interference terms while taking into account time-frequency resolution problem are also addressed.