Ioan Pavaloi

Emotion recognition in audio records

A novel approach to find the combination of acoustic features producing a more robust automatic recognition of a speaker emotion is proposed. Four discrete emotional states are classified. The emotional speech corpora used for training and evaluation are described in detail. The emotion recognition model using acoustic features is presented. The results achieved are presented and discussed.

Iris classification using WinICC and LAB color features

We present the WinICC software package, designed to help in tasks like clusterization or classification of images based on different feature vectors. The capabilities of this software are proven on a classification test involving 1.205 already segmented iris images belonging to 241 persons (five iris images per person - part of the UBIRISv1 Internet available database). We used the k-NN feature of the WinICC applied on LAB color feature vectors with 80 components extracted from iris images. The resulted rates of correctly classified irises are over 88% if 3 or 4 images are used to classify the remaining images of the same person. As the data set is not perfect, this is a result that may suggest a possible identification of human irises based on color distribution.

Automatic fury recognition in audio records

The paper is focused on automatic detection of fury emotion in audio records, using data extracted from the vocalic analysis of formants. We have studied speech prosody and voice inflexions and we recognised fury using classification algorithms applied to two databases, one with professional voices and another with normal voices, both of them recorded on the base of selected texts in Romanian language. We used relevant stories for generating fury emotion. We obtained interesting results that can be used in a large variety of possible applications.

Color feature vectors based on optimal LAB histogram bins

In this paper we propose an automatic method for computing the bin boundaries of complex 3D LAB histograms in order to extract optimal color feature vectors from digital images. The size of the feature vectors can be adapted to particular application needs. We tested our approach with very good results on an iris recognition problem solved empirically before.

Acoustic analysis methodology on Romanian language vowels for different emotional states

A methodology and an application for acoustic analysis of Romanian language vowels, for different emotional states and different vowel location in words, are presented. The application can be useful for expressivity analysis, speech synthesis and the speech to text approach. We focus on the data structures inside the software for annotation analysis of speech files. The formant statistical characteristic analysis of vowels on file subclasses selected by user is performed. The application allows refining the emotional state analysis at the level of vowels.

A study on automatic recognition of positive and negative emotions in speech

The paper is focused on an experimental study on positive and negative emotion vocal recognition. After some considerations about the positive and negative emotions, the paper gives a short description of the three corpuses used in the work we have accomplished. The paper describes three sets of coefficients used, the statistic features used to generate the three sets of feature vectors and the two classification methods used in this study. The recognition results obtained for every corpus are shown and some conclusions and directions of development are presented.

Experimental study in development of speech corpus for emotion recognition with data validation

The work on emotion recognition and models evaluation requires large corpora with recordings of emotional voices. The objective of this paper is to show a simple technique of automatic data validation that can be used in the development of a speech corpus. The paper describes an experimental study for a speech corpus development using two collections of data for vocal emotion expression with three emotions, happiness, anger, sadness and a normal (unemotional) state. In the validation step we used two classifiers, k-NN (k - Nearest Neighborhood) and SVM (Support Vector Machines), and five different sets of feature vectors based on formants F0-F4 values, MFCC (Mel-Frequency Cepstral Coefficients) and PLP (Perceptual Linear Prediction) coefficients values of the speech recording. The presented method is verified by human validation process in building an emotional recognition corpus.

Experiments on image classification and retrieval using statistics on pixels position

In this paper we have proposed a color indexing scheme for image classification and retrieval using color features. Experiments were made on the Corel 1000 database, for three different color spaces, LAB, HSV and RGB. In our tests, for image classification, two discriminative classifiers, k-NN (k - Nearest Neighborhood) and SVM (Support Vector Machine) were used. Two new distances were defined and used in k-NN experiments and the results were compared with results obtained using k-NN with three well known distances, Canberra, Euclidian and Manhattan. For image retrieval, the performance of the proposed method, measured in terms of average precision and average recall were compared with performance obtained with methods using color, texture and shape features. Our approach retrieves the highest number of relevant images compared to other more computationally expensive techniques. This color indexing method can improve the robustness of finding images with a similar color composition and can be used as a simple, fast and computationally simple filter, whose output can be then processed by other methods.

Iris recognition using statistics on pixel position

In this paper we present experiments made on an extension of a well-known simple global color criterion and we present the results obtained in iris recognition on two known iris databases, UBIRIS and UPOL. All tests were done for three different color spaces, RGB, HSV and LAB, and two discriminative classifiers, k-NN (k - Nearest Neighborhood) and SVM (Support Vector Machine) were used in our experiments. For the k-NN, three classic distances, Canberra, Euclidean and Manhattan and other two new distances we proposed were used in experiments. The results obtained using the extension color criterion are compared with the ones obtained with the original criterion.

Experimental study in emotion recognition using prosodie features

The paper describes an experimental study on emotion recognition using a collection of emotional recordings from SRoL corpus. Its goal is to study and to obtain a simple tool that can be used in recordings validation in the process of building large voice corpora. The tools can help or even replace the human validation. In this study we used two classifiers, k-NN (k - Nearest Neighborhood) and SVM (Support Vector Machines), and seven different sets of feature vectors based on F0 formant values. After a small introduction, the database and the feature vectors are presented. Commonly used evaluation measures, including Recall, Precision, Accuracy, MCC (Matthews correlation coefficient), computed from the confusion matrix obtained in emotion recognition are then enumerated. Next, there are presented and discussed the achieved results, that can be used in a large variety of possible applications. The presented work is validated by the human validation already done on SRoL corpus.