Halima Bahi

Combination of vector quantization and hidden Markov models for Arabic speech recognition

We present experiments performed to recognize isolated Arabic words. Our recognition system is based on a combination of the vector quantization technique at the acoustic level and Markovian modelling. Hidden Markov models (HMMs) are widely used in a number of practical applications and are especially suitable in speech recognition because of their ability to handle variability of the speech signal. In our system, a word is analysed and represented as a set of acoustic vectors, then transformed into a symbolic sequence using the vector quantizer. This observation sequence is compared to reference Markov models. The word associated with the model obtaining the highest score is declared to be the recognized word.

A hybrid approach for Arabic speech recognition

Summary form only given. A recent innovation in artificial research is the integration of multiple artificial intelligent techniques into hybrid intelligent systems. One of the most used integration is the neuro-symbolic one. Some research in this area deals with the integration of expert systems and neural networks. In particular, we are interested with the connectionist expert system (CES) introduced by Gallant; it consists of an expert system implemented throughout a multilayer perceptron. In such a network each neuron has a symbolic significance. We present a CES dedicated to the Arabic speech recognition. So, we implemented a neural network where the input layer represents the acoustical level, the hidden layer, the phonetic level, and the output layer stands for the lexical one.

A new keyword spotting approach

Keyword spotting is the task of identifying the occurrences of certain desired keywords in an arbitrary speech signal. Keyword spotting has many applications one of them is telephone routing. In particular, we consider a big company which receives thousands of telephone calls daily. We are interested with the classification of these calls to route them to the appropriate department. State-of-the-art approaches in keyword spotting considered the speech signal as continuous speech, where keywords and non-keywords are modeled using hidden Markov models. In this paper, we suggest a perceptual approach, which exploit acoustic particularities of the predefined keywords to detect their frontiers, without modeling the Out-Of-Vocabulary (OOV) words.

Type-2 Fuzzy Gaussian mixture models for singing voice classification in commercial music production

The paper describes a system of singing voice classification in the commercial music productions. A first step in our system is to separate the singers voice from the music. Based on the vocal part, two sets of parameters are formed, one for singing voice type and the other for the singing voice quality. Each set of parameters contains a number of MPEG–7 low–level descriptors and other descriptors; at the classification stage the paper suggests an extension of Gaussian Mixture Models (GMMs), by using the Type–2 FGMMs (Type–2 Fuzzy Gaussian Mixture Models). Results show substantial improvements when compared to similar works.

Embedded Learning Segmentation Approach for Arabic Speech Recognition

Building an Automatic Speech Recognition (ASR) system requires a well segmented and labeled speech corpus (often transcription is made by an expert). These resources are not always available for languages such as Arabic. This paper presents a system for automatic Arabic speech segmentation for speech recognition purpose. State-of-the-art models in ASR systems are the Hidden Markov Models (HMM), so that for the segmentation, we expect the use of embedded learning approach where an alignment between speech segments and HMMs is done iteratively to refine the segmentation. This approach needs the use of transcribed and labelled data, for this purpose, we built a dedicated corpus. Finally, the obtained results are close to those described in the literature and could be improved by handling more Arabic speech specificities.

Text extraction from historical document images by the combination of several thresholding techniques

This paper presents a new technique for the binarization of historical document images characterized by deteriorations and damages making their automatic processing difficult at several levels. The proposed method is based on hybrid thresholding combining the advantages of global and local methods and on the mixture of several binarization techniques. Two stages have been included. In the first stage, global thresholding is applied on the entire image and two different thresholds are determined from which the most of image pixels are classified into foreground or background. In the second stage, the remaining pixels are assigned to foreground or background classes based on local analysis. In this stage, several local thresholding methods are combined and the final binary value of each remaining pixel is chosen as the most probable one. The proposed technique has been tested on a large collection of standard and synthetic documents and compared with well-known methods using standard measures and was shown to be more powerful.

Using multi decision tree technique to improving decision tree classifier

The automatic classification systems, prediction and data mining are used in many applications marketing, finance, customer relationship management... using large databases. In this paper we describe a new data mining approach based on decision trees. In the proposed approach we built a multi-layer decision tree model, where each layer consists of several decision trees. The aim of the multi decision tree MDT is to improve decision tree classifier. The performances of MDT are compared with C4.5 decision tree algorithm and some ensemble of decision tree classifiers, namely bagging decision tree, boosting decision trees BDT and random forests decision tree. Results show substantial improvements when compared to these approaches.

Use of Gaussian Mixture Models and Vector quantization for singing voice classification in commercial music productions

Instead of the expansion of the information retrieval systems, the music information retrieval domain is still an open one. In this context, the singing voice classification is a promised trend. In this paper, we shall present our experiments concerning the classification of singers according to their voice type, and their voice quality. Some experiments were carried in which two sets of parameters are used in addition to the use of two classification approaches: The GMM (Gaussian Mixture Models) and the VQ (Vector quantization). The obtained results were compared to those provided by the related state-of-the-art approaches.

HMM/SVM segmentation and labelling of Arabic speech for speech recognition applications

Building a large vocabulary continuous speech recognition (LVCSR) system requires a lot of hours of segmented and labelled speech data. Arabic language, as many other low-resourced languages, lacks such data, but the use of automatic segmentation proved to be a good alternative to make these resources available. In this paper, we suggest the combination of hidden Markov models (HMMs) and support vector machines (SVMs) to segment and to label the speech waveform into phoneme units. HMMs generate the sequence of phonemes and their frontiers; the SVM refines the frontiers and corrects the labels. The obtained segmented and labelled units may serve as a training set for speech recognition applications. The HMM/SVM segmentation algorithm is assessed using both the hit rate and the word error rate (WER); the resulting scores were compared to those provided by the manual segmentation and to those provided by the well-known embedded learning algorithm. The results show that the speech recognizer built upon the HMM/SVM segmentation outperforms in terms of WER the one built upon the embedded learning segmentation of about 0.05%, even in noisy background.

A statistical-based decision for arabic pronunciation assessment

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.