Loading...

Speaker Recognition based on Telephone Speech

Khaki, Hossein | 2012

553 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 43346 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Ghaemmaghami, Shahrokh; Demiroğlu, Cenk
  7. Abstract:
  8. The speech signal renders different level of information from phoneme to identity even emotional information. By this specific information one can recognize the identity. In this thesis, multi-session speaker authentication with minimal speech data is addressed. Incomplete data and variable channel are the two most important problems that affect accuracy of speaker authentication systems. The goal of this project is to compensate the effects of channel and improve the error rate of the system in incomplete data conditions. First, the recent method on base of identity vectors is studied. For this issue the Mel Frequency Cepestrum Coefficient extracted and modeled with Gaussian Mixture Models which are adapted by Universal Background Models. Then Identity vectors are produced by total variability space method. Then the effective discriminant methods are applied to reduce the effect of channel and incomplete data. We Among these methods, Generalized Discriminant Analysis (GDA) as a kernel method is applied for two modes, 2 classes as a scoring method and multiclasses as a channel compensation method. In channel compensation mode, Probabilistic Linear Discriminant Analysis (PLDA) works better than the GDA. However, GDA can outperform in scoring mode with more than 98% accuracy. The simulation was done on the NIST 2006 and 2010 databases with 10 second and 5 minutes test data
  9. Keywords:
  10. Speaker Identification ; Generalized Discriminant Analysis ; Probabilistic Linear Discriminat Analysis ; Identity Vector (I-Vector)

 Digital Object List

 Bookmark

No TOC