Meeting Sehedule 2017


    What do you want to do?  
bult Log in to My Account
bult New Registration
bult GISFI Webmail
bult GISFI Newsletter
bult Image Gallery
bult Change Password
bult Benefits of joining GISFI
News & Events
more BACK

Ms. Swati Prasad has successfully defended his PhD thesis titled, "Efficient and Robust Speaker Identification System under Adverse Conditions" on Wednesday, September 12, 2018





Speaker identification system (or voice based biometric), which determines the speaker of a given speech utterance from a group of people, is predicted to have huge importance in near future. Speech being the natural means of communication for humans, fits well in our requirement for less complex and more secure means of identification, followed by interaction with the electronic devices, than the existing password-based methods. But speaker identification system faces poor performance due to a mismatch between the train and the test speech conditions. A mismatch can arise due to factors like differences in handset, transmission channel or microphone; environmental noise; emotional state of the person; voice disguise, etc. It is also referred to as mismatched problem. In this thesis, mismatched problem, which arises due to environmental noise and voice disguise has been focused. To tackle performance degradation because of environmental noise mismatch, a hybrid method for feature frame selection has been developed. It combines voice activity detection (VAD) and variable frame rate (VFR) analysis methods. The hybrid method efficiently captures the speech part rejecting the non-speech, and the changes in the temporal characteristics of the speech signal considering the signal to noise ratio.

Mismatched problem which arises because of the adoption of voice disguise is less researched and poses threat to the existing speaker identification systems. Therefore, mismatch due to voice disguise has been focused in this thesis. Robust methods have been developed at the feature, training and the testing level. It has been found that, the use of fixed frame shift, typically of 10ms, leading to a fixed frame rate for acquiring the frames for feature extraction, might not give the best identification accuracy under voice disguise. Therefore, a multiple-model framework which combines features obtained by utilizing three different frame shifts of 3ms, 6ms and 9ms has been developed, and it has shown improved performance over the fixed frame shift method. Four multistyle training strategies has been investigated for tackling voice disguise mismatch seen in a security conscious organization. It has shown encouraging results over the single style training. Further a fusion framework, utilizing the best two investigated multistyle training strategies has been proposed. It has shown an overall improved performance over single style training, investigated multistyle trainings and the multiple-model methods.

Finally, a method combining multiple frame rates for feature extraction and reliable frame selection at the decision level has also been developed. It has shown an overall better performance over the baseline methods. Since, voice production is governed by brain, an attempt has also been made to study the brain signal response for motor imagery tasks. This might prove beneficial in future for improving the speaker identification system. A Brain Computer Interface (BCI) to classify motor imagery tasks from the same limb has been studied. For this, a feature selection strategy for brain signal has also been proposed. It consisted of channel selection based on fisher ratio and time-segment selection by visual inspection. It has shown improved performance over the baseline system.

Assessment Committee:
Associate Professor Sofoklis Kyriazakosh, Aarhus University (Chair)
Professor Seshadri Mohan, University of Arkansas af Little Rock. Arkansas, United States
Professor Gregory Yovanof, Athens Information Technology (AIT), Athens Greece.

Professor Ramjee Prasad, Aalborg University, Denmark