Skip to Main Content
The usual methods for speech signal analysis give a spectrum evaluation on a time interval (20 to 40 ms) larger than the largest pitch period, which eliminates most phonetically relevant information. Our approach consists in a decomposition of the signal into a string of "impulses" defined in the frequency-time domain. The work described in the present paper is done using a filterbank followed by short-time integrators. An "Impulse Coherence Function" is defined ; its maxima are used to mark the places where the impulse are. Classification experiments are carried out, in order to segment the speech wave according to the regularity and similarity of the successive impulses. Applications may be envisioned in the fields of voicing and pitch detection, consonant features recognition, and auditory modeling.