A new method for automatic definition of emotion dependent features is proposed. We proposed the ratio of emotional expression similarity between two speakers and carried out the multilingual analysis of emotional speech. The analysis revealed that almost all speakers in the analysis used unique sets of emotion-dependent features and each speaker used unique values for the defined emotion-dependent features. This achievement present the fundamental definition of procedure for automatic extraction of emotion dependent features.
COBISS.SI-ID: 10374422
A new method for automatic data-driven generation of phonetic broad classes was proposed. The phonetic broad classes are needed for initialization of decision tree, which is being used for clustering of acoustic models for speech recognition. The method is very suitable for multi-lingual case, where it is difficult to find an expert with equal knowledge of all languages in the pool. Such knowledge is needed in the phase of initializing multi-lingual decision trees for developing automatic speech recognition system. The new data-driven approach improved the speech recognition results.
COBISS.SI-ID: 10087958
A digital parametric filter was proposed, which is capable of identifying skin color in real time. The proposed solution enables the development of procedures for dynamic adoption of filter parameters according to the user’s skin color. The developed system combines computer generated stereoscopic graphics with information collected from stereoscopic camera and projects such visual information on stereoscopic data eyeglasses. The proposed system was patented as type of spatial visual-gesture communication.
COBISS.SI-ID: 9671446
A new propositional branching-time temporal logic ACTLW (Action-based Computation Tree Logic with unless operator) has been developed. It is intended for formal specification of properties of systems described by using event-based models. Algorithms for global model checking of ACTLW formulae by using fixed-point calculations and symbolic representation have been developed. They have been implemented in the proprietary software package EST (Efficient Symbolic Tools). The logic and the algorithms have already been tested by automatic verification of many reactive systems with EST.
COBISS.SI-ID: 12047638
A novel computationally efficient noise robust feature extraction algorithm using combined wavelet packet decomposition and autoregressive modeling of speech signals was developed. The performance was compared to the standardized feature extraction procedure ETSI ES 201 108. With the proposed method, relative automatic speech recognition improvements of 44.75 % and 48.26 % were achieved. The procedure can be applied in speech recognition systems that operate in real-life environments since it enables strong reduction of additive noises thus increasing the overall robustness of such systems.
COBISS.SI-ID: 11170838