Dorra Ben Ayed

On the use of different feature extraction methods for linear and non linear kernels

The speech feature extraction has been a key focus in robust speech recognition research; it significantly affects the recognition performance. In this paper, we first study a set of different feature extraction methods such as linear predictive coding (LPC), mel frequency cepstral coefficient (MFCC) and perceptual linear prediction (PLP) with several features normalization techniques including rasta filtering and cepstral mean subtraction (CMS). Based on this, a comparative evaluation of these features is performed on the task of text independent speaker identification using a combination between gaussian mixture models (GMM) and linear or non-linear kernels based on support vector machine (SVM).

Improved Frame Level Features and SVM Supervectors Approach for The Recogniton of Emotional States from Speech: Application to Categorical and Dimensional States

The purpose of speech emotion recognition system is to classify speakers utterances into different emotional states such as disgust, boredom, sadness, neutral and happiness. Speech features that are commonly used in speech emotion recognition (SER) rely on global utterance level prosodic features. In our work, we evaluate the impact of frame -level feature extraction. The speech samples are fro m Berlin emot ional database and the features extracted fro m these utterances are energy, different variant of mel frequency cepstrum coefficients (MFCC), velocity and accelerat ion features. The idea is to explo re the successful approach in the literature of speaker recognition GMM-UBM to handle with emotion identification tasks. In addition, we propose a classification scheme for the labeling of emotions on a continuous dimensional-based approach. Index Terms—speech emotion recognition, valence, arousal, MFCC, GMM Supervector, SVM

Adaboost with SVM using GMM supervector for imbalanced phoneme data

In machine learning, AdaBoost with Support vector Machines (SVM) based component classifier have shown to be a successful method for classification on balanced dataset with all classes having relatively similar distribution. However, the success of this method is limited when it is applied for imbalanced datasets. In many real applications, the classification of data with imbalanced proportions will be problematic since the algorithm can be biased and then might predict all the samples into majority classes. Many studies were conducted to overcome imbalance data problem by using hybrid algorithms. In this paper, we propose an improved AdaBoost with SVM based weak learner algorithm using Gaussian Mixture Modeling (GMM) supervectors called GSV-ADSVM. GMM supervectors are constructed applying MAP adaptation of the means of the mixture components based on speech from a target phoneme of TIMIT corpus. Those supervectors will be used as input datasets for the hybrid Adaboost-SVM. The main goal of this paper is to investigate the impact of using GMM supervectors with the boosted SVM in a multi-class phoneme recognition problem with the aim to advance the classification of imbalanced data since certain class of interest have very small size.

Cooperative static and dynamic neural networks for phoneme recognition

This paper investigates the use of static and dynamic neural networks in phoneme recognition. Besides this, the paper also proposes a cooperative static and dynamic neural networks model. The cooperative model integrates a decision system for phoneme recognition. Mel cepstrum coding has been applied to represent speech signal in frames. Features from the selected frames are used to train neural networks based models. The comparative study show that the proposed cooperative model provides more accurate recognition rates both in auto-coherence test and generalization test.

Incorporating Belief Function in SVM for Phoneme Recognition

The Support Vector Machine (SVM) method has been widely used in numerous classification tasks. The main idea of this algorithm is based on the principle of the margin maximization to find an hyperplane which separates the data into two different classes.In this paper, SVM is applied to phoneme recognition task. However, in many real-world problems, each phoneme in the data set for recognition problems may differ in the degree of significance due to noise, inaccuracies, or abnormal characteristics; All those problems can lead to the inaccuracies in the prediction phase. Unfortunately, the standard formulation of SVM does not take into account all those problems and, in particular, the variation in the speech input. This paper presents a new formulation of SVM (B-SVM) that attributes to each phoneme a confidence degree computed based on its geometric position in the space. Then, this degree is used in order to strengthen the class membership of the tested phoneme. Hence, we introduce a reformulation of the standard SVM that incorporates the degree of belief. Experimental performance on TIMIT database shows the effectiveness of the proposed method B-SVM on a phoneme recognition problem.

Robust Noisy Speech Recognition Using Deep Neural Support Vector Machines

This paper aims to classify noisy sound samples in several daily indoor and outdoor acoustic scenes using an optimized deep neural networks (DNNs). The advantage of a traditional DNNs lies in using at the top layer a softmax activation function which is a logistic regression in order to learn the output label in a multi-class recognition problem. In this paper, we optimize the DNNs by replacing the softmax activation function by a linear support vector machine.

Feature Selection Using Adaboost for Phoneme Recognition

The propose to improve a Support Vector Machines (SVM) learning accuracy by using a Real Adaboost algorithm for selecting features is presented in the chapter. This technique aims to minimize the recognition error rates and the computational effort. Hence, the Real Adaboost will be used not as classifier but as a technique for selecting features in order to keep only the relevant features that will be used to improve our systems accuracy. Since the Real Adaboost is only used for binary classifications problems, we investigate different ways of combining selected features applied to a multi-class classification task. To experiment this selection, we use the phoneme datasets from TIMIT corpus [Massachusetts Institute of Technology (MIT), SRI International and Texas Instruments, Inc. (TI)] and Mel-Frequency Cepstral Coefficients (MFCC) feature representations. It must be pointed out that before using the Real Adaboost the multi-class phoneme recognition problem should be converted into a binary one.

The challenges of SVM optimization using Adaboost on a phoneme recognition problem

The use of digital technology is growing at a very fast pace which led to the emergence of systems based on the cognitive infocommunications. The expansion of this sector impose the use of combining methods in order to ensure the robustness in cognitive systems.