Event

Practice Talk for Qual Exam

Howard Lei

ICSI

Tuesday, August 26, 2008
12:30

In this work, I propose to investigate structured approaches to data selection for speaker recognition, with an emphasis on information theoretic approaches. As a side effect, I will attempt to understand why certain speaker recognition systems perform better than others based on their data usage, and conversely, why certain data allow for better speaker recognition systems. Once this is completed, I will attempt to use this knowledge to implement an effective data selection procedure, and implement a speaker recognition system optimized for the data used. One upshot of this is that I may be able to predict how well a speaker recognition system in its blue prints will behave (including its fusion with other systems) without actually implementing the system. A second upshot is that this will provide for a better approach to data reduction for speaker recognition, allowing for faster modeling, and easier data storage.

In particular, I will be dealing primarily with HMM and GMM based speaker recognition systems, which comprise the vast majority of current state-of-the-art speaker recognition systems. I will investigate ways to effectively select data for training and testing, involving various scientific measures. I also plan to investigate the performances of certain common and off-the-shelf SVMspeech-lunch@icsi kernel techniques as they relate to the measures used to select the data. I will investigate methods to make these measures easily obtainable, such that the amount of work required to extract these measures from data will pale in comparison to the work required to run an entire speaker recognition system.