Halima Bahi

Speech Disorders Recognition using Speech Analysis

Multi-Modal Advancements

INTRODUCTION When a person is unable to produce speech sounds correctly or fluently, or has probl... more INTRODUCTION When a person is unable to produce speech sounds correctly or fluently, or has problems with his or her voice, then he/she has a speech disorder. Difficulties pronouncing sounds, or articulation disorders, and stuttering are examples of speech disorders. When a person has trouble understanding others (receptive language), or sharing thoughts, ideas, and feelings completely (expressive language), then he/she has a language disorder. A stroke can result in aphasia, or a language disorder. Both children and adults can have speech and language disorders. They can occur as a result of a medical problem or have no known cause.

Download

An Arabic Mispronunciation Detection System by means of Automatic Speech Recognition Technology

by Khaled Necibi and Halima Bahi

The aim of the work presented in this paper is to help the Algerian (Arabic) young learners havin... more The aim of the work presented in this paper is to help the Algerian (Arabic) young learners having problems with pronunciation of Arabic language to improve their pronunciation skills through the use of ASR technology (Automatic Speech Recognition). The proposed system is based on mispronunciation detection methods where different scores of pronunciation are calculated to decide how the word was badly pronounced in term of quantitative measure. The obtained results have showed that the GLL (Global average Log Likelihood) score detects mispronunciation with 86.66% of correct rejection. To validate this finding, another experiment was carried out, on an expanding corpus of data, based on measuring the calculated mispronunciation scores in terms of CA (Correct Acceptance), CR (Correct Rejection), FA (False Acceptance) and FR (False Rejection). The obtained results have shown that the GLL score, comparing to the other calculated scores, had the higher CA+CR (76.66%) and the lower FA+FR (23.32%).

A statistical-based decision for arabic pronunciation assessment

by Khaled Necibi and Halima Bahi

The aim of a computer assisted language learning (CALL) system is to improve the language skills ... more The aim of a computer assisted language learning (CALL) system is to improve the language skills of learners. Such systems often include, grammar and vocabulary components, while the pronunciation learning seems to be the hardest step in language learning process. Little attention has been paid to this aspect among the required ones in CALL systems. In pronunciation learning context, the learner would like to know if its pronunciation is good or bad. In the case where the pronunciation is bad, it will be suitable if some advices are given to him. The goal of this work is an early detection of pupils with reading difficulties and in the issue of decision whether their pronunciation is good or not is our particular interest. For this purpose, we consider the answer to this question as a classification problem and we use a statistical approach to make a decision; this approach allows us to pursue the investigation concerning the pronunciation of every phoneme in the word or in the sentence.

Download

Feature Embedding Representation for Unsupervised Speaker Diarization in Telephone Calls

Communications in computer and information science, Nov 4, 2023

Anomaly detection with a variational autoencoder for Arabic mispronunciation detection

International journal of speech technology, Jun 25, 2024

Computer-assisted language learning (CALL) systems increasingly arouse a significant interest and... more Computer-assisted language learning (CALL) systems increasingly arouse a significant interest and establish a presence in automated foreign language learning. They enhance traditional learning methods by providing access to various accents and spoken language styles through websites, mobile applications, and social media. Herein, mispronunciation detection is a key component mainly addressed as a classification problem. Meanwhile, deep learning (DL) advances have promoted these systems by training deep neural networks (DNN) to classify a pronunciation as correct or incorrect. However, the effectiveness of the DL models is hindered by many shortcomings, such as the scarcity of labeled data. To address this issue, the paper assumes an anomaly detection-based mispronunciation detection approach. It utilizes a variational autoencoder (VAE) relying on a density-based method to model the “normal data.” The VAE is a generative model trained in a self-supervised way to learn the distribution of the correct pronunciations, standing for “normal data,” and is expected to detect mispronunciations, standing for “abnormal data” during the test stage. Our proposition was evaluated in the context of Arabic pronunciation learning through the ASMDD Arabic dataset. The obtained results are promising, with an accuracy of about 98%. The proposed VAE outperformed the standard autoencoder as well as the state-of-the-art convolution neural networks used for Arabic mispronunciation detection.

Mispronunciation detection and diagnosis using deep neural networks: a systematic review

Multimedia tools and applications, Jan 9, 2024

The increased need for foreign language learning, along with advances in speech technology have h... more The increased need for foreign language learning, along with advances in speech technology have heightened interest in computer-assisted pronunciation teaching (CAPT) applications. Herein, the automatic diagnosis of pronunciation errors is essential, it allows language learners to identify their mispronunciations and thus improve their oral skills. Meanwhile, the emergence of deep learning algorithms for speech processing led to the use of deep neural networks at several stages of the mispronunciation detection and diagnosis process. Therefore, an overview of the state-of-the-art in deep learning algorithms for mispronunciation diagnosis is needed, for which we performed a systematic literature review. This study aims to provide an overview of the recent use of deep neural networks for mispronunciation detection and diagnosis (MDD). A thorough statistical analysis is provided in this review which was conducted by extracting specific information from 53 papers published between the years 2015 and 2023. This review indicates that the diagnosis of pronunciation errors is a highly active area of research. Quite a few deep learning models and approaches have been proposed in this area, but there are still some important open issues and limitations to be addressed in future works.

Self-Supervised Spoken Term Detection for Query by Example

Ingénierie des systèmes d'information/Ingénierie des systèmes d'Information, Jun 20, 2024

Technological advances in signal processing and enormous data storage capabilities have made larg... more Technological advances in signal processing and enormous data storage capabilities have made large speech datasets readily available. These speech data often being unlabelled and unannotated their effective use is a research hot topic. In particular, query by example spoken term detection (QbE-STD) constitutes a potential way to take advantage of these resources. Given a speech stream, QbE-STD techniques aim to detect the presence of a given spoken query, in an unsupervised way. QbE-STD is considered a pattern-matching problem, and the dynamic time-warping (DTW) algorithm is state-of-the-art. This paper proposes an alternative to the DTW and its variants by leveraging deep learning models. The proposed architecture uses convolutions to extract short-term dependencies and an attention mechanism to extract long-term dependencies, from the speech archive. Thus, a Fully Connected Convolutional Neural Network (FC-CNN) is trained to detect the presence of a target spoken query in a speech stream. During the training stage, the model is trained on raw speech, in a self-supervised way, while during the test stage and without prior knowledge on the target query, the system is expected to decide whether the query exists or not within the speech archive. Experiments were carried on a dataset built on the Google speech commands corpus. The obtained results show the potential of our suggestion in indexing speech databases.

Download

Anomaly detection with a variational autoencoder for Arabic mispronunciation detection

International Jouranal of Speech Technology, 2024

Computer-assisted language learning (CALL) systems increasingly arouse a significant interest and... more Computer-assisted language learning (CALL) systems increasingly arouse a significant interest and establish a presence in automated foreign language learning. They enhance traditional learning methods by providing access to various accents and spoken language styles through websites, mobile applications, and social media. Herein, mispronunciation detection is a key component mainly addressed as a classification problem. Meanwhile, deep learning (DL) advances have promoted these systems by training deep neural networks (DNN) to classify a pronunciation as correct or incorrect. However, the effectiveness of the DL models is hindered by many shortcomings, such as the scarcity of labeled data. To address this issue, the paper assumes an anomaly detection-based mispronunciation detection approach. It utilizes a variational autoencoder (VAE) relying on a density-based method to model the “normal data.” The VAE is a generative model trained in a self-supervised way to learn the distribution of the correct pronunciations, standing for “normal data,” and is expected to detect mispronunciations, standing for “abnormal data” during the test stage. Our proposition was evaluated in the context of Arabic pronunciation learning through the ASMDD Arabic dataset. The obtained results are promising, with an accuracy of about 98%. The proposed VAE outperformed the standard autoencoder as well as the state-of-the-art convolution neural networks used for Arabic mispronunciation detection.

Mispronunciation detection and diagnosis using deep neural networks: a systematic review

Multimedia Tools and Applications, 2024

The increased need for foreign language learning, along with advances in speech technology have h... more The increased need for foreign language learning, along with advances in speech technology have heightened interest in computer-assisted pronunciation teaching (CAPT) applications. Herein, the automatic diagnosis of pronunciation errors is essential, it allows language
learners to identify their mispronunciations and thus improve their oral skills. Meanwhile, the emergence of deep learning algorithms for speech processing led to the use of deep neural networks at several stages of the mispronunciation detection and diagnosis process. Therefore, an overview of the state-of-the-art in deep learning algorithms for mispronunciation diagnosis is needed, for which we performed a systematic literature review. This study aims to provide an overview of the recent use of deep neural networks for mispronunciation detection and diagnosis (MDD). A thorough statistical analysis is provided in this review which was
conducted by extracting specific information from 53 papers published between the years 2015 and 2023. This review indicates that the diagnosis of pronunciation errors is a highly active area of research. Quite a few deep learning models and approaches have been proposed in this area, but there are still some important open issues and limitations to be addressed in future works.

Self-Supervised Spoken Term Detection for Query by Example

Ingénierie des Systèmes d’Information, 2024

Technological advances in signal processing and enormous data storage capabilities have made larg... more Technological advances in signal processing and enormous data storage capabilities have made large speech datasets readily available. These speech data often being unlabelled and unannotated their effective use is a research hot topic. In particular, query by example spoken term detection (QbE-STD) constitutes a potential way to take advantage of these resources. Given a speech stream, QbE-STD techniques aim to detect the presence of a given spoken query, in an unsupervised way. QbE-STD is considered a pattern-matching problem, and the dynamic time-warping (DTW) algorithm is state-of-the-art. This paper proposes an alternative to the DTW and its variants by leveraging deep learning models. The proposed architecture uses convolutions to extract short-term dependencies and an attention mechanism to extract long-term dependencies, from the speech archive. Thus, a Fully Connected Convolutional Neural Network (FC-CNN) is trained to detect the presence of a target spoken query in a speech stream. During the training stage, the model is trained on raw speech, in a self-supervised way, while during the test stage and without prior knowledge on the target query, the system is expected to decide whether the query exists or not within the speech archive. Experiments were carried on a dataset built on the Google speech commands corpus. The obtained results show the potential of our suggestion in indexing speech databases.

Download

Software connector as software process

International Journal of Simulation and Process Modelling, 2021

International audienc

Comparative Study of CNN Structures for Arabic Speech Recognition

Ingénierie Des Systèmes D'information, Apr 30, 2023

Speech recognition is an essential ability of human beings and is crucial for communication. Cons... more Speech recognition is an essential ability of human beings and is crucial for communication. Consequently, automatic speech recognition (ASR) is a major area of research that is increasingly using artificial intelligence techniques to replicate this human ability. Among these techniques, deep learning (DL) models attract much attention, in particular, convolutional neural networks (CNN) which are known due to their power to model spatial relationships. In this article, three CNN architectures that performed well in recognized competitions were implemented to compare their performance in Arabic speech recognition; these are the well-known models AlexNet, ResNet, and GoogLeNet. These models were compared based on a corpus composed of Arabic spoken digits collected from various sources, including messaging and social media applications, in addition to an online corpus. The architectures of AlexNet, ResNet, and GoogLeNet achieved respectively an accuracy of 86.19%, 83.46%, and 89.61%. The results show the superiority of GoogLeNet, and underline the potential of CNN architectures to model acoustic features of low-resource languages such as Arabic.

Download

One-Class Training for Intrusion Detection

Proceedings of the 1st International Conference on Intelligent Systems and Pattern Recognition, 2020

The extensive use of the internet has generated a large amount of data, and a new class of threat... more The extensive use of the internet has generated a large amount of data, and a new class of threats has arisen which led to the emergence of a new field of computer science: intrusion detection. An Intrusion Detection System (IDS) scans the data generated by the network traffic to detect potential attacks. In this paper, we propose a method for anomaly detection, in the network traffic, based on deep learning algorithms using a new approach called training objective. We propose to use a Deep Auto-Encoder (DAE) with One-Class training to analyze the contents of a dataset; as the normal accesses are more likely to be recognized, we utilize normal examples to train the DAE. Meanwhile, the dropout is applied to deal with the overfitting problem. We report experimental results regarding the accuracy, recall, and precision of the system on the well-known NSL-KDD benchmark dataset for network intrusion. Overall, the results indicate that the DAE model trained using One-class examples positi...

AlgerianChidrenCAPT_1_4

The contains pronunciations from nine pupils aging from 5 to 8 years; each of them uttered a set ... more

Combination of a DAE-CNN and OC-SVDD for intrusion detection

International Journal of Computational Systems Engineering, 2021

Remote Spoken Arabic Digits Recognition Using CNN

Lecture Notes in Electrical Engineering, 2022

While Arabic is the fifth most spoken language in the world, it still lacks powerful speech recog... more While Arabic is the fifth most spoken language in the world, it still lacks powerful speech recognition systems. Moreover, well-resourced languages have benefited greatly from recent developments in artificial intelligence and In deep learning, while the development of Arabic speech recognizers is still modest. In this paper, we contribute to bridging the gap between the resources of Arabic and those of the well-resourced languages, in particular in the context of remote communications. Since the unavailability of dedicated corpora is the most problematic that prevents the development of Arabic products, we built a corpus composed of the ten Arabic digits. The corpus was collected remotely, over-connected smartphone apps, and social media. A convolutional neural network algorithm is then trained to recognize the digits. Initial experiments faced an overfitting problem due to insufficient data. To overcome this problem, we artificially expanded the training dataset with audio data augmentation techniques. In further experiments, we explored cross-corpora speech recognition. The final model achieved a recognition rate of 95.94% on training subset and 77.30% on test subset. The results showed the potential of the proposed approach in remote speech recognition conditions.

Download

Song classification

Abstract—We live in a world full of data. Every day, people handle different types of data from a... more Abstract—We live in a world full of data. Every day, people handle different types of data from a variety of measurements and observations. Data describes the characteristics of a species of life, depicts the properties of a natural phenomenon; summarizes the results of experiences or records the dynamics of system operation. Thus, the data provides a basis for analysis, reasoning, decision and ultimately, for understanding all kinds of objects and phenomena. The new information technology, at the same time they facilitate the movement and storage of information, contributing to their exponential growth, they are in numbers disproportionate to the human resources to address them. This paper addresses the problem of indexing database containing songs using title and artist particularities to enable their effective exploitation. In this article, we present our experiments in automated song categorization, where we suggest the use of an ant colony algorithm. A Naive Bayes algorithm is ...

Download

Tree Network Data mining approach

The aim of this research is to propose ‎data ‎mining approach based on decision trees. In this ‎a... more The aim of this research is to propose ‎data ‎mining approach based on decision trees. In this ‎approach we ‎build a ‎model as a network that contains several ‎layers, in ‎each ‎layer we have several decision ‎trees, and then we ‎propose a method for using our model. ‎The objective of DTN is to ‎improve the present existing data mining‎‎techniques. So, we compare the Decision Tree ‎Network (DTN) ‎performances with some well-known ‎methods, namely Boosting Decision Trees (BDT) and ‎bagging ‎decision tree; we also compare DTN with C4.5 ‎decision tree algorithm. Results show ‎substantial improvements when compared to similar techniques‎.‎ MOTS-CLÉS : arbre de décision, fouille de données, algorithmes des arbres de décision, techniques de fouille de données.

Download

A New System For Audio Mining: Using of MDT+ for automatic prediction of new singers success

Proceedings of the 7th International Conference on Software Engineering and New Technologies, 2018

Online music diffusion and distribution is becoming increasingly important and commercial and per... more Online music diffusion and distribution is becoming increasingly important and commercial and personal databases are increasing in a considerable way. Nowadays, it's necessary to have tools that allow classifying and reaching these bases by carrying out analyzes through musical contents. The work included in this paper consists of automatic prediction of new singers success. This work lies within the scope of audio mining including of the heterogeneous data. Work can be divided into two parts: in the first part, we treat the audio data to make a classification of the Algerian singers' voices. There exist two categories of the singing voices classification (type and quality), each category has several classes. Within this framework, we proposed to use a vector of characteristics which contains the descriptors of MPEG-7 + the descriptors Not-MPEG-7 and Standard T2 FGMMs "Type 2 Fuzzy Gaussian Mixture Models" for modeling and classification of singing voice. By using ...

Data mining approach

The aim of this research is to propose ‎‎approach based on decision trees. In this ‎we ‎a ‎as a n... more The aim of this research is to propose ‎‎approach based on decision trees. In this ‎we ‎a ‎as a network that contains several ‎in ‎‎we have several decision ‎and then we ‎a method for using our model. ‎objective of DTN is to ‎ the present existing data mining‎‎ So, we compare the Decision Tree ‎ (DTN) ‎ with some well-known ‎namely Boosting Decision Trees (BDT) and ‎‎tree; we also compare DTN with C4.5 ‎tree algorithm. Results show ‎improvements when compared to similar techniques‎‎ MOTS-CLES : arbre de decision, fouille de donnees, algorithmes des arbres de decision,

Uploads

Papers by Halima Bahi

Log In