Analysis of Event-Related Potentials Elicited During Speaker Recognition

Analysis of Event-Related Potentials Elicited During Speaker Recognition Tasks Dror Lederman 1, Joseph Tabrikian 1, Hillel Pratt 2 (1) Ben-Gurion University of the Negev and (2) Technion- Israel Institute of Technology Research questions: • What information concerning identification of different speakers can be extracted from the EEG measurements? Is it possible to perform speaker identification (in the general sense) based on EEG measurements? • Which brain areas are responsible for identification of speakers? • What is the spatio-temporal course of the physiological processes associated with speaker identification? Model-based approaches for ERP estimation Multi-channel ERP Classification 1. HMM-based ERP averaging Problems/tasks: • Decomposition of the ERPs from the noisy EEG measurements. • Multi-channel classification of the ERP signals. • Estimation and localization of the ERPs' intra-cranial sources. Primary research objective: Develop statistical model-based (GMM/CD-HMMs) methods for ERPs estimation and classification and utilize these methods to answer the major research questions. Preliminary classification results 2. Parallel Model Combination Method Secondary research objective: Investigate the spatio-temporal course of the processes that occur in the brain during speaker recognition tasks based on ERPs source localization (LORETA). Research plan • Acquisition of speaker identification AERPs databases. Three databases will be acquired: v Speaker gender identification AERPs. v Familiar/unfamiliar speaker identification AERPs. v Speaker identification AERPs. • ERP estimation methods development. The methods will be evaluated using: v Synthesized DB. v AERPs speaker recognition DB. • ERP classification methods development and evaluation. • Spatio-temporal course investigation (LORETA). Classification of two ERP classes elicited during imagery left/right hand movements tasks (2 nd BCI competition) Features Classification Error Rate (%) Left-To-Right CD-HMM with 5 states and 3 Gaussian/state (the current work) Bayes classifier (Christin and Schafer) Linear discriminant analysis (Akash and Narayana) Let the vectors , and (each of size ) represent the ongoing EEG, the ERP and the noisy EEG measurements, respectively, such that. Since the zero-mean background EEG is presumed to be WSS, its covariance matrix, , , is Toeplitz. 0. 7 Morlet-wavelets at 10 and 22 Hz in channels C 3 and C 4 Ratio of AR spectral power in 4 frequency bands, channels C 3 and C 4 10. 7 Several neural networks trained on different time regions (Saffari) 3. Stationary-nonstationary decomposition 10 Filter bank coefficients extracted from the first channel of single-trial ERPs AAR parameters (no details were given) 15. 7 13. 6 Preliminary work conclusions üThe CD-HMM-based approach outperforms other acceptable approaches. üThe research has the potential to contribute to the improvement of ERPs processing methods in general, to provide additional insight into the speaker recognition cognitive process and take the BCI research one step forward. ü The approaches presented here utilize a-priori statistical information on the background EEG and ERP signals and the statistical properties of these signals. Consequently, they are expected to outperform current ERP estimation methods. Therefore: The solutions can be found using Lagrange-multipliers. This method is an extension to the Minimum Mean Square Error approach with the EEG stationarity constraints. In memory of Prof. Arnon Cohen who initiated this project Prof. ARNON COHEN 5. 02. 1938 - 19. 02. 2005