Present position: Temporary researcher
|Thesis title:||Separazione di sorgenti acustiche tramite informazioni geometriche dell'ambiente riverberante|
This thesis concerns the semi-blind separation of acoustic sources in reverberant environments using microphone arrays. A novel methodology which performs separation
in two steps is presented. First, the localization of the sources is achieved through the analysis of their Time Differences of Arrival (TDOAs) at different microphone
pairs. Two ways can be followed to estimate TDOAs: the estimation of the channels between each microphone and each source or the estimation of the de-mixing filters. In this thesis we propose to use a Blind Source Separation approach called TRINICON  to localize sources. It was demonstrated that TRINICON is effective in mildly reverberating environments. However, when reverberation time exceeds a threshold or sources are moving, de-mixing filters cannot track the variations of the channels. In Chapter 3 several experiments demonstrate that localization can be achieved although the separation degree is not acceptable. Once the sources are localized, in the second step, a beamformer is designed to suppress the jammer. Moreover, when we acquire signals in a reverberant environment, a significative portion of the incoming energy comes from reflections. As a consequence, the beamformer
design stage must keep into account also the suppression of the echoes of the interferer.
Given the environment geometry and the position of sources and microphones, a fast beam tracing algorithm compute the Directions of Arrival of the reflections.
A test is proposed to assess which reflections are significative for the design of the beamformer.
Several experiments are shown to illustrate the overall performance of the proposed separation algorithm. In particular experiments conducted in a real-case scenario
show, as expected, the strict dependency between localization accuracy and Signal to Interferer Ratio.