Adam Kurowski

Separability Assessment of Selected Types of Vehicle-Associated Noise

Music Information Retrieval (MIR) area as well as development of speech and environmental information recognition techniques brought various tools intended for recognizing low-level features of acoustic signals based on a set of calculated parameters. In this study, the MIRtoolbox MATLAB tool, designed for music parameter extraction, is used to obtain a vector of parameters to check whether they are suitable for separation of selected types of vehicle-associated noise, i.e.: car, truck and motorcycle. Then, cross-correlation between pairs of parameters is calculated. Parameters for which absolute value of cross-correlation factor is below a selected threshold, are chosen for further analysis. Subsequently, pairs of parameters found in the previous step are analyzed as a graph of low-correlated parameters with the use of the Bron-Kerbosch algorithm. Graph is checked for existence of cliques of parameters linked in all-to-all manner related to their low correlation. The largest clique of low-correlated parameters is then tested for suitability for separation into three vehicle noise classes. Behrens-Fisher statistic is used for this purpose. Results are visualized in the form of 2D and 3D scatter plots.

Speech Analytics Based on Machine Learning

In this chapter, the process of speech data preparation for machine learning is discussed in detail. Examples of speech analytics methods applied to phonemes and allophones are shown. Further, an approach to automatic phoneme recognition involving optimized parametrization and a classifier belonging to machine learning algorithms is discussed. Feature vectors are built on the basis of descriptors coming from the music information retrieval (MIR) domain. Then, phoneme classification beyond the typically used techniques is extended towards exploring Deep Neural Networks (DNNs). This is done by combining Convolutional Neural Networks (CNNs) with audio data converted to the time-frequency space domain (i.e. spectrograms) and then exported as images. In this way a two-dimensional representation of speech feature space is employed. When preparing the phoneme dataset for CNNs, zero padding and interpolation techniques are used. The obtained results show an improvement in classification accuracy in the case of allophones of the phoneme /l/, when CNNs coupled with spectrogram representation are employed. Contrarily, in the case of vowel classification, the results are better for the approach based on pre-selected features and a conventional machine learning algorithm.

Assessment of Therapeutic Progress After Acquired Brain Injury Employing Electroencephalography and Autoencoder Neural Networks

A method developed for parametrization of EEG signals gathered from participants with acquired brain injuries is shown. Signals were recorded during therapeutic session consisting of a series of computer assisted exercises. Data acquisition was performed in a neurorehabilitation center located in Poland. The presented method may be used for comparing the performance of subjects with acquired brain injuries (ABI) who are involved in concentration training program. It may also allow for an assessment of relative difference in performance of two participants involved to exercises by comparing parameters derived from EEG signals acquired in the course of therapeutic sessions. The parametrization method is based on autoencoder neural networks. The efficiency of parameters extracted employing the algorithm was compared to parameters derived from the spectrum of EEG signal. As it was confirmed by achieved results, the presented autoencoder-based method may be applied to predict ABI subjects’ performance in attention training sessions.

Analysis of results of large-scale multimodal biometric identity verification experiment

An analysis of a large set of biometric data obtained during the enrolment and the verification phase in an experimental biometric system installed in bank branches is presented. Subjective opinions of bank clients and of bank tellers were also surveyed concerning the studied biometric methods in order to discover and to explore relations emerging from the obtained multimodal dataset. First, data acquisition and identity verification methods are described in this study. Then, relationships between ratios of successful and failed verifications between pairs, triplets, and quartets of biometric modalities are studied. An analysis of the sentiment of clients and of banking tellers related to each identity verification attempt was performed based on linguistic methods. The data mining process is described, based on the rough sets methodology, aimed at deriving rules pertaining to consecutive identity verification attempts.

Performance Analysis of Developed Multimodal Biometric Identity Verification System

The bank client identity verification system developed in the course of the IDENT project is presented. The total number of five biometric modalities including: dynamic handwritten signature proofing, voice recognition, face image verification, face contour extraction and hand blood vessels distribution comparison have been developed and studied. The experimental data were acquired employing multiple biometric sensors installed at engineered biometric terminals. The biometric portraits of more than 10 000 bank clients were registered and stored in the database during the presented study and then verified experimentally. Problem- specific survey was done on the basis of questionnaires completed by the subjects in order to assess the look and feel of the developed biometric system as well as to collect opinions concerning its implementation in banking outlets. A discussion concerning the quality of registered data and results achieved in the study is included.

Comparison of selected electroencephalographic signal classification methods

A variety of methods exists for electroencephalographic (EEG) signals classification. In this paper, we briefly review selected methods developed for such a purpose. First, a short description of the EEG signal characteristics is provided. Then, a comparison between the selected EEG signal classification methods, based on the overview of research studies on this topic, is presented. Examples of methods included in the study are: Artificial Neural Networks, Support Vector Machines, Fuzzy or k-Means Clustering. Similarities and differences between all considered methods of an automatic EEG signal classification with a focus on consecutive stages of such a process are reviewed. Examples of EEG classification, considering various types of usage and target applications along with their effectiveness, are also shown.

Multimodal System for Diagnosis and Polysensory Stimulation of Subjects with Communication Disorders

An experimental multimodal system, designed for polysensory diagnosis and stimulation of persons with impaired communication skills or even non-communicative subjects is presented. The user interface includes an eye tracking device and the EEG monitoring of the subject. Furthermore, the system consists of a device for objective hearing testing and an autostereoscopic projection system designed to stimulate subjects through their immersion in a virtual environment. Data analysis methods are described, and experiments associated with classification of mental states during listening exercises as well as audio-visual stimuli are presented and discussed. Feature extraction was based on discrete wavelet transformation and clustering employing the k-means algorithm was designed. All algorithms were implemented in the Python programming language with the use of Open Source libraries. Tests of the proposed system were performed in a Special School and Educational Center in Kościerzyna, Poland. Results and comparison with data gathered from the control group of healthy people are presented and discussed.