print

Design and Implementation of Neural Network Algorithms in Cepstrum-Based Speaker Identification for SGU Students Log In

---

Identification by using speech data, being only one type biometric feature measurements, is by no means a new method, but the prospective applications in this field are undoubtedly enormous.nIt is the objective of the thesis to give guidelines in designing such a Speaker Identification System. As the problem of identifying a speaker based on his/her speech characteristics is a pattern recognition problem, Artificial Neural Network is one of the obvious choices for solving this type of non-linear problem.nSome may argue that the Artificial Neural Network advantage is also its downfall: it offers solutions without a means of explaining the reasoning process behind the solutions, hence an Artificial Neural Network is usually looked upon as a 'black box' solution.nLike in most problems, there are no exact one way to solve a Neural Network problem. Various approaches are then studied and applied so that the designing time can be greatly reduced, and people can concentrate on employing the model into a live system.nThe Speaker Identification system is based on the assumption of a small number of users, i.e. not more than 10, inspired by Furui's research [5] where he achieved an error of 0.2% at 3-second-long utterances with pattern matching techniques.nFurthermore, the thesis then focuses on the use of short information (word per word or a segment of a word) instead of long information usually associated with Speech Recognition. This is based on the fact that one's speech features are individual to even a small window span (for example 256 samples window-size at ~30 msec). nA Multi-Layer Perceptron is then utilized and the result is a system that achieves a 0.3% recognition error for trained data, 5% recognition error for untrained data, and 2.5% rejection error for speakers outside the database.nThe thesis, while focusing on student log-in area, can be utilized into different applications, as the process of feature extraction and network designing are mostly similar.

---

Availability

M00033

(wh)

Available

Detail Information

Series Title

-

Call Number

33

Publisher

: Swiss German University., 2003

Collation

-

Language

English

ISBN/ISSN

-

Classification

-

Content Type

-

Media Type

-

Carrier Type

-

Edition

-

Subject(s)

MIT

Specific Detail Info

-

Statement of Responsibility

-

Other version/related

File Attachment

Comments

---

You must be logged in to post a comment