Détection des occlusives à l'aide de la transformée en ondelettes

Abstract
This paper describes the detection of stop consonants in the French language with the help of the wavelet transform. Our hypothesis is that the burst of stop consonants can be approximated with a pulse. The detection system is based on the study of the correlation functions between the modulus of the impulse wavelet transform and the same transform of the speech signal. As we are using a Gaussian wavelet, the correlation functions have a minimum before each maximum, where that maximum is synchronous with the burst of the stop consonant. Two signals S_M and S_m are computed. They are respectively the synchronous and asynchronous sums of all the correlation functions. To localize the local minima of the S_m function, two criteria are used. The temporal localisation of the pulse is accomplished with certain local maxima of the signal S_M. The method of detection is tested with a corpus which has 137 unvoiced and 55 voiced stop consonants. The rate of detection is 94% for the unvoiced stops and 75% for the voiced stops. This study shows that for detection, the burst of stop consonants can be modelled with a pulse.