TY  - THES
T1  - Adaptive time-frequency analysis for cognitive source separation
A1  - Kümmel,Sylvia
Y1  - 2010/08/02
N2  - This thesis introduces a framework for separating two speech sources in non-ideal, reverberant environments. The source separation architecture tries to mimic the extraordinary abilities of the human auditory system when performing source separation. A movable human dummy head residing in a normal office room is used to model the conditions humans experience when listening to complex auditory scenes. This thesis first investigates how the orthogonality of speech sources in the time-frequency domain drops with different reverberation times of the environment and shows that separation schemes based on ideal binary time-frequency-masks are suitable to perform source separation also under humanoid reverberant conditions. Prior to separating the sources, the movable human dummy head analyzes the auditory scene and estimates the positions of the sources and the fundamental frequency tracks. The source localization is implemented using an iterative approach based on the interaural time differences between the two ears and achieves a localization blur of less than three degrees in the azimuth plane. The source separation architecture implemented in this thesis extracts the orthogonal timefrequency points of the speech mixtures. It combines the positive features of the STFT with the positive features of the cochleagram representation. The overall goal of the source separation is to find the ideal STFT-mask. The core source separation process however is based on the analysis of the corresponding region in an additionally computed cochleagram, which shows more reliable Interaural Time Difference (ITD) estimations that are used for separation. Several algorithms based on the ITD and the fundamental frequency of the target source are evaluated for their source separation capabilities. To enhance the separation capabilities of the single algorithms, the results of the different algorithms are combined to compute a final estimate. In this way SIR gains of approximately 30 dB for two source scenarios are achieved. For three source scenarios SIR gains of up to 16 dB are attained. Compared to the standard binaural signal processing approaches like DUET and Fixed Beamforming the presented approach achieves up to 29 dB SIR gain.
KW  - Akustik
KW  - Hören
KW  - Orthogonalität
CY  - Saarbrücken
PB  - Universitäts- und Landesbibliothek
AD  - Postfach 151141, 66041 Saarbrücken
UR  - http://scidok.sulb.uni-saarland.de/volltexte/2010/3245
ER  -