Creating and sharing knowledge for telecommunications

Vocalic Filler Characterization using Acoustical Properties Derived from Cepstrum

Veiga, A. ; Proença , J. ; Celorico, D. ; Candeias, S. ; Perdigão, F.

Vocalic Filler Characterization using Acoustical Properties Derived from Cepstrum, Proc Jornadas en Tecnología del Habla and III Iberian SLTech Workshop - IberSPEECH , Madrid, Spain, Vol. -, pp. 260 - 268, November, 2012.

Digital Object Identifier:

Download Full text PDF ( 130 KBs)

Abstract
The study presented here attempts to characterize the so-called vocalic fillers, according to their acoustic and prosodic properties. A comparison with the vowels of the phonemic system of the European Portuguese possessing similar timbre is also provided. Among the acoustic and prosodic parameters used to characterize such speech segments, in this contribution we paid particular attention to the spectral shape (timbre) and duration. Such parameter analysis was conducted by means of training HMM models, as is usually done for speech recognition purposes, from which a spectrum is derived. This is an extension of a previous work towards the characterization of the hesitation phenomena with a new methodology. The corpus used is a subset of the Portuguese television broadcast news (RTP) and of the Tecnovoz speech database. This study allowed us to characterize two of the most common vocalic fillers in Portuguese spontaneous speech as well as the two unstressed vowels with similar timbre.