Spotting Arabic Phonetic Features Using Modular Connectionist Architectures and a Rule-Based System.
Sid‐Ahmed Selouani, Jean Caelen · 1998
This paper reports the results of experiments in complex Arabic phonetic features identification using a rulebased system (SARPH) and modular connectionist architectures. The first technique we present, operates in the field of analytic approaches and intends to implement a relevant system for automatic segmentation and labeling through the use of finite state networks (FSN). For this task, an original ear model is used to calculate indicative features according to the phonetic and phonological matrix of standard Arabic we have established in earlier studies. The second method deals with a set of a simplified version of sub-neural-networks (SNN). A binary sub-task is assigned to these networks with the objective to recognize features as subtle as emphasis, gemination and semantically pertinent lengthening of vowels. This is proposed to be done at two different levels: at the gross level, by identifying the macro-classes, at the finer level, by detecting pertinent temporal distortion and emphasis. Serial and parallel architectures of SNN are investigated. A comparison between the two identification strategies is carried out using stimuli uttered by Algerian native speakers. The results show that SNN achieved well in rough identification while in the some cases of phonologic duration the rule-based system performs better. 1.