Hocine Teffahi’s research while affiliated with University of Sciences and Technology Houari Boumediene and other places

Automatic Speech Recognition (ASR) technology plays an essential role in human-machine interaction. In this paper, we describe our conducted speech experiments that are realized to develop and adapt a mixed automaticspeech recognition system based on spoken digits for Amazigh and Arabic dialects that are considered as under-resourced dialects in the Maghreb region.Our used database includes speech samples collected from 24 Moroccans and Algerian speakers including both males and females. The designed system is implemented based on the combination of hidden Markov models and Gaussian mixture models, as well as the Mel frequency spectral coefficients (MFCCs) feature extraction method.

This paper investigates a transfer learning approach to solve the spoken dialects identification problem for some under-resourced dialects of the Maghrebi region, including Algerian Arabic Dialect (AAD), Algerian Berber Dialect (ABD), Moroccan Arabic Dialect (MAD), andMoroccan Berber Dialect (MBD). In our experiments, we used different Transfer learning models, namely: Residual Neural Network (Resnet50, Resnet101), and Visual Geometric Group (VGG16, VGG19) using anin-house corpus that we built for each dialect. The corpus is composed of ten digits recorded for each of the aforementioned dialects, repeated ten times by six native speakers. The results vary according to different reasons: the number of epochs, neurons, batch size, and also the datasets combinations used in training and test phases. The best score found is 90.4% by the VGG19 model. Overall, the results show the robustness of our system based on the VGG16 model with an average identification rate of 62.7%.

Par ses propriétés morphologiques et syntaxiques la langue arabe est considérée comme une langue très difficile à maîtriser dans le domaine du traitement automatique et les systèmes de synthèse à partir du texte arabe sont donc très peu nombreux. Le but de notre travail est une réalisation d'un système de synthèse de la parole par concaténation (sélection) dynamique d'unités dans un corpus pour la langue arabe baptisé GArabic TTS sous l'environnement Matlab. Le texte à introduire est un texte non voyellé qui facilite l'utilisation du système, la sortie est disponible uniquement pour une voix masculine. L'évaluation de GArabic TTS est basée sur un test subjectif et objectif ; en ce qui concerne l'intelligibilité, aspects naturels (l'écoute) et la qualité (PESQ). L'évaluation finale de la qualité globale du système est jugée satisfaite.

In this paper, we present a set of experiments aiming to improve the recognition of spoken digits for under-resourced dialects of the Maghrebi region, using a hybrid system. Indeed, integrating a Dialect Identification module into an Automatic Speech Recognition (ASR) system has shown its efficiency in previous works. In order to make the ASR system able to recognize digits spoken in different dialects, we trained our hybrid system on Moroccan Berber Dialect “MBD,” Moroccan Arabic Dialect “MAD,” and Algerian Arabic dialect “AAD,” in addition to Modern Standard Arabic. We have investigated five machine learning based classifiers and two deep learning models: the first one is based on Convolutional Neural Network (CNN), and the second one uses two pre-trained models: Residual Deep Neural Network (Resnet50 and Resnet101). The findings show that the CNN model outperforms the other proposed methods and consequently enhances the performance of spoken digit recognition system by 20% for both Algerian and Moroccan dialects.

The general objective of this paper is to build a system in order to automatically recognize emotion in speech. The linguistic material used is a corpus of Arabic expressive sentences phonetically balanced. The dependence of the system on speaker is an encountered problem in this field; in this work we will study the influence of this phenomenon on our result. The targeted emotions are joy, sadness, anger and neutral. After an analytical study of a large number of speech acoustic parameters, we chose the cepstral parameters, their first and second derivatives, the Shimmer, the Jitter and the duration of the sentence. A classifier based on a multilayer perceptron neural network to recognize emotion on the basis of the chosen feature vector that has been developed. The recognition rate could reach more than 98% in the case of an intra-speaker classification and 54.75% in inter-speaker classification. We can see the system’s dependence on speaker clearly.

The absence of the diacritical marks from the modern Arabic text generates a significant increase of the ambiguity in the Arabic text, which can cause confusion in the pronunciation of a written word. Despite the fact that the reader with a certain level of Arabic knowledge can easily recover the missing diacritics by: using the words context, the morphology and the syntax knowledge of the Arabic language. This paper describes a design and implementation of a Text-To-Speech system for a diacritic Arabic text. The goal of this project is to obtain a set of high quality speech synthesizer based on unit selection using a bi-grams model taking into account the particularities of the language. It takes a diacritic Arabic text as input and produces corresponding speech; the output is available as male voice. The evaluation of our TTS system is based on subjective and objective tests. The final evaluation of GArabic TTS system, regarding the intelligibility, naturalness aspects (listening) and the quality (PESQ) is jugged successful.

In this article we present the methodology employed for the design and evaluation of a Basic Arabic Expressive Speech corpus (BAES-DB). The corpus, which has a total length of approximately 150 minutes, is constituted of 13 speakers uttering a set of 10 sentences while simulating 3 emotions (joy, anger and sadness) in addition to a neutral utterance. The 10 sentences have been selected to meet phonetic equilibrium and absence of emotional content criteria, from a corpus of sentences proposed in Boudraa [1]. The corpus was evaluated through various tests of guided categorization, performed through a website. The overall good recognition rate is currently at 83.03%, with sadness being the most well recognized type of expressive speech (90.93%) and joy being the least well recognized (73.87%).

In this paper we will present a contribution to the design of an expressive speech synthesis system for the Arabic language. The system uses diphone concatenation as the synthesis method for the generation of 10 phonetically balanced sentences in Arabic. Rules for the orthographic-to-phonetic transcription are detailed, as well as the methodology employed for recording the diphone database. The sentences were synthesized with both “neutral” and “sadness” expressions and rated by 10 listeners, and the results of the test are provided.

This paper deals with the underlining and the prominent display of the compensatory strategies developed by six speakers which are from different geographic regions when they express themselves in a second language very different from their mother native language. The paradigm that had been used in this study is the elocution speed as a speaking constraint in order to investigate the effects of the flow rate on the acoustical parameters of the speech signal. The analysis is performed in both time and frequency domains. We have obtained very interesting results as we have found out that the used constraint allows a good discrimination between different speakers on the basis of their native language.