Localization and separation of concurrent talkers based on principles of auditory scene analysis and multi-dimensional statistical methods

This work focuses on 'statistical cocktail party processing': localization, tracking, and enhancement of voices in concurrent speech or in high levels of nonstationary noise. Key strategies are to link binaural and spectro-temporal information, to combine cues across frequency and time by a probabilistic approach, and to treat speech as a multidimensional stochastic signal, using a priori knowledge about it. To implement these, Bayesian estimation, sequential Monte Carlo methods, and statistical evaluation of speech databases are used. Three on-line algorithms are developed and tested, which run partly in real-time. They allow for a robust, efficient and exact sound localization even at low signal-to-noise ratios (SNRs), and successful tracking and separation of voices with convergence times between 50 and 200 ms. The multidimensional statistical approach allows to analyze acoustical scenes at low SNR, showing that the described strategies might help to interpret auditory processing.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Oldenburg

Year

1970

Author dc:creator

• Nix, Johannes

• [Keine Schlagwörter von Autor/in vergeben.]

Repository record source_url

http://oops.uni-oldenburg.de/89

OAI identifier oai:identifier

oai:oops.uni-oldenburg.de:89