SIGNAL AND DATA PROCESSING
Year:2009 | Volume:- | Issue:2 (SERIAL 10)|Papers Count:7

In this paper, a model of spoken word recognition is proposed. This model is particularly concerned with extraction of cues from the signal leading to a specification of a word in terms of bundles of distinctive features, which are assumed to be the building blocks of words. In the model proposed, auditory input is chunked into a set of successive time slices. It is assumed that the derivation of the underlying word pattern proceeds in three layers: Features, phonemes, words. The feature layer has a complete set of feature detectors at every time slice. In this layer, the detection of the underlying pattern of distinctive features from the speech signal proceeds 'in three steps. In the first step, numerical values for features are obtained measuring acoustic attributes in each time slice. The acoustic attributes are either acoustic landmarks corresponding to articulator-free features which are identified, based on amplitude changes in various energy bands, or acoustic cues in the vicinity of the landmarks corresponding to articulator-bound features. Continuous perceptual feature values are, then processed into a much more structured representation, namely phonological surface structure. This is carried out in Perception Grammar as suggested by Boersma (1998). In the third step, a further processing is carried out. to turn the discrete representation into an abstract one yielding the underlying pattern of distinctive features. The next layer of the model has a complete set of phoneme detectors for every three time slices, but each set spans six time slices so the sets overlap. This means that the detection of adjacent phonemes will also overlap; this is supposed to simulate coarticulation The top layer has a complete set of word detector centered on every three time slices; again, the sets overlap, the number of time slices per word detector is variable because it depends on the length of each individual word.

The object of this research is development of memory assessment system, using Event Related Potentials. Our approach is using ERPs recorded on Fz, Cz and pz electrodes. Subjects made old/new recognition judgments on new unstudied unmeaning pictures and old pictures which had been presented at study phase. Features related with memory activity in time-frequency domain were used to achieve this purpose. So that discrete wavelet transform coefficients of Event Related Potentials computed, then using mean, variance and power of specific frequency bands, 36 features on 3 channels were obtained entirely. After appropriate feature selection on single and three channels Linear discriminant analysis was done to get classification results using selected features. Finally we discriminated groups with 89.3% accuracy in test group by combination of three channels features.

The autonomic nervous system controlled the cardiovascular system by two antagonistic parts of the autonomic nervous system (ANS), which was named as the cardiac sympathetic and the parasympathetic nervous. Human emotion changing is related to sympathetic nervous and parasympathetic nervous system. Since the dynamics of the cardiac system are nonlinear. Nonlinear methods have been applied to the analysis of HRV. We have included studies that used the three most common non-linear indices: the fractal scaling exponent, measures based on Poincare plots, wavelet entropy, and common linear indices in time and frequency domain. These results suggest that non-linear analysis contains relevant information related to different HRVdynamics in normal and abnormal subjects.

Many important speech enhancement methods operate in frequency domain. In these methods, a frequency gain filter is multiplied by the noisy spectrum. In this paper, frequency methods are investigated and the gain filters are shown as functions of the Signal to Noise Ratio (SNR). Since gain filters are functions of SNR, SNR estimation methods are presented and consequently a new SNR estimation method is proposed. Moreover, we proposed a technique for controlling the level of residual noise and shaping it. In our proposed techniques, the frequency gain filters are tuned in order to have less noise reduction and as a result achieve less distortion. In addition to controlling the level of noise, in our proposed technique, the shape of residual noise is changed in order to have more pleasant residual noise. For evaluation of methods, listening tests are performed and the results are reported based of four aspects of speech signals: musical residual noise, level of background noise, echo and unnatural residual noise. Evaluation of methods shows that the proposed methods are successful from these aspects point of view.

In this study we propose a new approach to analyze data from the P300 speller paradigm using the quadratic B-Spline wavelet coefficients in comparing to time and frequency features sets on the event related potentials. Data set II from the BCI competition 2005 was used. Mode frequency, Mean frequency, Median frequency and some morphologic parameters ware extracted as features. Three methods were used for comparing three feature subsets, first Davies Bouldin criteria, correlation based method' and classification accuracy criteria. For all criteria, best result was extracted from wavelet coefficients, at the final wavelet coefficients were used as inputs into committee machines (CM) based on LDA, MLP and SVM. This algorithm achieved an accuracy of 97.6% for train data and 94.2% for test data of subject A in target and non-target detection also accuracy of 98.2% for train data and 92.8% for test data of subject B.

An embolus is a blood clot, a fat globule or gas bubbles that may be freely circulating in bloodstream can stop the blood flow and lead to ischemia. In real time assessment of blood flow by Trans Cranial Doppler (TCD) method, travelling solid or gaseous micro emboli in the bloodstream by passing across the assessment area, causes a short time signal with high intensity. While TCD recording including movement of the probe, coughing, sneezing, and head rotation It generate high intensity artifacts that make it difficult to make differentiate from embolus. Time consuming and also human mistakes in differentiating emboli from artifact are the main motivations of design of the automatic detection systems. Implementing such systems is nowadays faced with two main challenges problems: extracting suitable features and designing the proper classifier. In this research, we studied two issues together. In feature extraction part, wavelet coefficient, wavelet entropy, fractal dimension and Besov property of signal is extracted, and using by statistical methods we introduced the feature with highest separability rate. In classifier part, a novel method based on hidden markov models for detecting emboli from artifact is proposed, and the results is compared with the results of Adaptive Neuro Fuzzy Inference System classifier. In total, using wavelet coefficients and hidden markov model, we achieved an accuracy rate of 95.3% and specificity of 92.7%.

In this paper Structural Gaussian Mixture Model (SGMM) method, which used for GMM-UBM algorithm speed-up in speaker verification system, investigated. Effects of some parameters in Structural Background Model (SBM) construction studied in detail and optimal values are used in- model creation. In addition a structure named "GMM identifier" proposed for processing output scores of SBM-SGMM structure. Simulation results demonstrate that SBM optimal structure combined with proposed score processing unit outperforms the base system with neural network while the proposed system complexity is less than neural network. Using proposed method a speed-up rate of 2.7 attained while system performance improved compared to GMM-UBM system. Best configuration leads to an Equal Error Rate of 0.35% having considerable improvement considering GMM-UBM system performance of 1.71% in Equal Error Rate prospective.