Arabic discourse analysis based on acoustic, prosodic and phonetic modeling: elocution evaluation, speech classification and pathological speech correction

This work describes a complete study addressing the pathological speech processing. It focuses principally on the speech correction and the assistance to learners of Arabic vocabulary. For this purpose, we follow five main phases. The first one is dedicated to evaluate the produced speech by assigning a pronunciation level for each speaker according to their forced alignment score. The second step consists in classifying the Arabic produced speech into healthy or pathological based on two different models: a prosodic modeling based on elocution speed and a phonetic modeling based on comparing between a referenced Probabilistic-Phonetic Model and a speaker model. Third, we localize for each speech sequence classified as pathological the problematic phonemes that degrade pronunciation. We differentiate also two factors which can falsify produced acoustic signals: degraded speech can be generated from pathological problems, or it can be produced by non arabophone pronouncers. Hence, we focus on forced alignment scores. Fourth, we develop a new algorithm to correct pathological pronunciation. We opt of two different solutions: lexical and phonetic. The last task is the conception of an application assisting learners of Arabic vocabulary to improve their pronunciation. The achieved results are encouraging. Moreover, the evaluation and classification of produced acoustic signals are satisfactory, learners of Arabic vocabulary have presented good amelioration using the developed application. A lot of applications that design systems of voice signal processing and platforms of e-learning can enjoy from our proposition.