Tunisian Dialect Recognition Based on Hybrid
Techniques
Mohamed Hassine, Lotfi Boussaid, and Hassani Massaoud
Laboratoire de Recherche ATSI, Ecole
Nationale d’Ingénieurs de Monastir, Tunisia
Abstract: In this research paper, an
Arabic Automatic Speech Recognition System is implemented in order to recognize
ten Arabic digits (from zero to nine) spoken in Tunisian dialect (Darija). This
system is divided in two main modules: The feature extraction module by
combining a few conventional feature extraction techniques, and the recognition
module by using Feed-Forward Back Propagation Neural Networks (FFBPNN). For
this purpose, four oral proper corpora are prepared by five speakers each. Each
speaker pronounced the ten digits five times. The chosen speakers are different
in gender, age and physiological conditions. We focus our experiments on a
speaker dependent system and we also examined the case of speaker independent
system. The obtained recognition performances are almost ideal and reached up
to 98.5% when we use for the feature extraction phase the Perceptual Linear
Prediction technique (PLP) followed firstly by its first-order temporal
derivative (∆PLP ) and secondly by Vector Quantization of Linde-Buzo-Gray
(VQLBG).
Keywords: Vector Quantization (VQLBG), Mel Frequency Cepstral Coefficients
(MFCCs), Feed-Forward Back Propagation Neural Networks
(FFBPNN),
Speaker Dependent System.
Received April 24, 2015; accept February 3, 2017