zvukové signály

Disertační práce se zabývá obecným problémem slepé separace zdrojů, zejména se zaměřuje na zvukové signály. Dodatek disertační práce pojednává o automatické klasifikaci biomedicínských dat.Práce se skládá ze tří hlavních částí. V rámci první části je popsán algoritmus pro analýzu nezávislých komponent (ICA) nazvaný Block EFICA. Algoritmus je určen pro separaci okamžikových směsí nestacionárních signálů. Je také ukázáno, že algoritmus je asymptoticky eficientní za předpokladu, že zdrojové signály jsou blokově stacionární a mají konstantní rozptyl.Druhá a třetí část se zabývají separací konvolutorních směsí zvukových zdrojů. Autor zde prezentuje modifikaci pro existující algoritmus T-ABCD, jehož základem je opět ICA. Modifikace předpokládá nedokonalou separaci provedenou ICA a na základě tohoto předpokladu zpřesňuje rekonstrukci zdrojů. Dále je navržena ?online? verze T-ABCD, která je schopna adaptovat separaci vzhledem ke změnám v procesu míchání a také vzhledem k nestacionaritě zdrojů.; The thesis deals with the general problem of Blind Source Separation (BSS), themain focus being on audio sources. In an appendix, a research outside the topic of BSS is disscused: The automatic classification of biomedical signals.The general goal of the blind source separation is to estimate unknown sourcesignals from their measured mixtures, while there is no information about the mixing process available. This task is too general and cannot be solved without additional assumptions about the unknown sources. One such assumption is the mutual independence of the source signals.Methods performing the separation based on the independence assumption are denoted as Independent Component Analysis (ICA).Most algorithms described in this thesis are based on ICA.The thesis consists of three main parts.The first part deals with the separationof non-stationary instantaneous mixtures. The author contributed to the proposal of an ICA algorithm called Block EFICA. This algorithm is proved to be asymptotically efficient (i.e. gives the most accurate estimates of the sources possible) provided that the source signals are block-wise stationary and have constant variances.The second and third part concerns the separation of convolutive mixtures ofaudio sources. This problem can be solved via ICA in the time domain, for example by an existing method known as T-ABCD, which was proposed Koldovskýand Tichavský. Modifications for this method are presented. These modificationsimprove the reconstruction of the sources in the case of imperfect ICA separation.Furthermore, an "online" version of the T-ABCD algorithm is proposed. It is ableto adapt to the changes in the mixing system and to the non-stationarity of thesource signals.The appendix deals with automatic classification of medical signals which originate in the screening examination of lower limbs arteries. This examination, performed via inexpensive ultrasound units, aims at an early diagnostics of PeripheralArterial Disease. Automatic classifiers for the data measured by mentioned ultrasound units are designed and trained. These classifiers are able to assign the signals to predefined classes, which reflect the degree of arterial occlusion.

Blind audio source separation via independent component analysis