Iosif Mporas

Affective speech interface in serious games for supporting therapy of mental disorders

We describe a novel design, implementation and evaluation of a speech interface, as part of a platform for the development of serious games. The speech interface consists of the speech recognition component and the emotion recognition from speech component. The speech interface relies on a platform designed and implemented to support the development of serious games, which supports cognitive-based treatment of patients with mental disorders. The implementation of the speech interface is based on the Olympus/RavenClaw framework. This framework has been extended for the needs of the specific serious games and the respective application domain, by integrating new components, such as emotion recognition from speech. The evaluation of the speech interface utilized purposely collected domain-specific dataset. The speech recognition experiments show that emotional speech moderately affects the performance of the speech interface. Furthermore, the emotion detectors demonstrated satisfying performance for the emotion states of interest, Anger and Boredom, and contributed towards successful modelling of the patients emotion status. The performance achieved for speech recognition and for the detection of the emotional states of interest was satisfactory. Recent evaluation of the serious games showed that the patients started to show new coping styles with negative emotions in normal stress life situations.

Speech segmentation using regression fusion of boundary predictions

In the present work we study the appropriateness of a number of linear and non-linear regression methods, employed on the task of speech segmentation, for combining multiple phonetic boundary predictions which are obtained through various segmentation engines. The proposed fusion schemes are independent of the implementation of the individual segmentation engines as well as from their number. In order to illustrate the practical significance of the proposed approach, we employ 112 speech segmentation engines based on hidden Markov models (HMMs), which differ in the setup of the HMMs and in the speech parameterization techniques they employ. Specifically we relied on sixteen different HMMs setups and on seven speech parameterization techniques, four of which are recent and their performance on the speech segmentation task have not been evaluated yet. In the evaluation experiments we contrast the performance of the proposed fusion schemes for phonetic boundary predictions against some recently reported methods. Throughout this comparison, on the established for the phonetic segmentation task TIMIT database, we demonstrate that the support vector regression scheme is capable of achieving more accurate predictions, when compared to other fusion schemes reported so far.

Automatic speech recognition performance in different room acoustic environments with and without dereverberation preprocessing

The performance of recent dereverberation methods for reverberant speech preprocessing prior to Automatic Speech Recognition (ASR) is compared for an extensive range of room and source-receiver configurations. It is shown that room acoustic parameters such as the clarity (C50) and the definition (D50) correlate well with the ASR results. When available, such room acoustic parameters can provide insight into reverberant speech ASR performance and potential improvement via dereverberation preprocessing. It is also shown that the application of a recent dereverberation method based on perceptual modelling can be used in the above context and achieve significant Phone Recognition (PR) improvement, especially under highly reverberant conditions.

Seizure detection using EEG and ECG signals for computer-based monitoring, analysis and management of epileptic patients

Seizure detection module using EEG and ECG signals as part of a eHealth system.Large scale evaluation of time and frequency domain features for seizure detection.Feature ranking evaluation and feature subsets performance evaluation.Operation in offline and online (real-time) mode. In this paper a seizure detector using EEG and ECG signals, as a module of a healthcare system, is presented. Specifically, the module is based on short-time analysis with time-domain and frequency-domain features and classification using support vector machines. The seizure detection module was evaluated on three subjects with diagnosed idiopathic generalized epilepsy manifested with absences. The achieved seizure detection accuracy was approximately 90% for all evaluated subjects. Feature ranking investigation and evaluation of the seizure detection module using subsets of features showed that the feature vector composed of approximately the 65%-best ranked parameters provides a good trade-off between computational demands and accuracy. This configurable architecture allows the seizure detection module to operate as part of a healthcare system in offline mode as well as in online mode, where real-time performance is needed.

Audio features selection for automatic height estimation from speech

Aiming at the automatic estimation of the height of a person from speech, we investigate the applicability of various subsets of speech features, which were formed on the basis of ranking the relevance and the individual quality of numerous audio features Specifically, based on the relevance ranking of the large set of openSMILE audio descriptors, we performed selection of subsets with different sizes and evaluated them on the height estimation task In brief, during the speech parameterization process, every input utterance is converted to a single feature vector, which consists of 6552 parameters Next, a subset of this feature vector is fed to a support vector machine (SVM)-based regression model, which aims at the straight estimation of the height of an unknown speaker The experimental evaluation performed on the TIMIT database demonstrated that: (i) the feature vector composed of the top-50 ranked parameters provides a good trade-off between computational demands and accuracy, and that (ii) the best accuracy, in terms of mean absolute error and root mean square error, is observed for the top-200 subset.

A hybrid architecture for automatic segmentation of speech waveforms

In the present work, we propose a hybrid architecture for automatic alignment of speech waveforms and their corresponding phone sequence. The proposed architecture does not exploit any phone boundary information. Our approach combines the efficiency of embedded training techniques and the high performance of isolated-unit training. Evaluating on the established for the task of phone segmentation TIMIT database, we achieved an accuracy of 83.56%, which corresponds to improving the baseline systems accuracy by 6.09 %.

Improving phone duration modelling using support vector regression fusion

In the present work, we propose a scheme for the fusion of different phone duration models, operating in parallel. Specifically, the predictions from a group of dissimilar and independent to each other individual duration models are fed to a machine learning algorithm, which reconciles and fuses the outputs of the individual models, yielding more precise phone duration predictions. The performance of the individual duration models and of the proposed fusion scheme is evaluated on the American-English KED TIMIT and on the Greek WCL-1 databases. On both databases, the SVR-based individual model demonstrates the lowest error rate. When compared to the second-best individual algorithm, a relative reduction of the mean absolute error (MAE) and the root mean square error (RMSE) by 5.5% and 3.7% on KED TIMIT, and 6.8% and 3.7% on WCL-1 is achieved. At the fusion stage, we evaluate the performance of 12 fusion techniques. The proposed fusion scheme, when implemented with SVR-based fusion, contributes to the improvement of the phone duration prediction accuracy over the one of the best individual model, by 1.9% and 2.0% in terms of relative reduction of the MAE and RMSE on KED TIMIT, and by 2.6% and 1.8% on the WCL-1 database.

Context-adaptive pre-processing scheme for robust speech recognition in fast-varying noise environment

Based on the observation that dissimilar speech enhancement algorithms perform differently for different types of interference and noise conditions, we propose a context-adaptive speech pre-processing scheme, which performs adaptive selection of the most advantageous speech enhancement algorithm for each condition. The selection process is based on an unsupervised clustering of the acoustic feature space and a subsequent mapping function that identifies the most appropriate speech enhancement channel for each audio input, corresponding to unknown environmental conditions. Experiments performed on the MoveOn motorcycle speech and noise database validate the practical value of the proposed scheme for speech enhancement and demonstrate a significant improvement in terms of speech recognition accuracy, when compared to the one of the best performing individual speech enhancement algorithm. This is expressed as accuracy gain of 3.3% in terms of word recognition rate. The advance offered in the present work reaches beyond the specifics of the present application, and can be beneficial to spoken interfaces operating in fast-varying noise environments.

Improving classification of epileptic and non-epileptic EEG events by feature selection

Correctly diagnosing generalized epileptic from non-epileptic episodes, such as psychogenic non epileptic seizures (PNES) and vasovagal or vasodepressor syncope (VVS), despite its importance for the administration of appropriate treatment, life improvement of the patient, and cost reduction for patient and healthcare system, is rarely tackled in the literature. Usually clinicians differentiate between generalized epileptic seizures and PNES based on clinical features and video-EEG. In this work, we investigate the use of machine learning techniques for automatic classification of generalized epileptic and non-epileptic events based only on multi-channel EEG data. For this purpose, we extract the signal patterns in the time domain and in the frequency domain and then combine all features across channels to characterize the spatio-temporal manifestation of seizures. Several classification algorithms are explored and evaluated on EEG epochs from 11 subjects in an inter-subject cross-validation setting. Due to large number of features feature ranking and selection is performed prior to classification using the ReliefF ranking algorithm within two different voting strategies. The classification models using feature subsets, achieved higher accuracy compared to the models using all features reaching 95% (Bayesian Network), 89% (Random Committee) and 87% (Random Forest) for binary classification (epileptic versus non-epileptic). The results demonstrate the competitiveness of this approach as opposed to previous methods.

Online Seizure Detection from EEG and ECG Signals for Monitoring of Epileptic Patients

In this article, we investigate the performance of a seizure detection module for online monitoring of epileptic patients. The module is using as input data streams from electroencephalographic and electrocardiographic recordings. The architecture of the module consists of time and frequency domain feature extraction followed by classification. Four classification algorithms were evaluated on three epileptic subjects. The best performance was achieved by the support vector machine algorithm, with more than 90% for two of the subjects and slightly lower than 90% for the third subject.