Page 1 |
Save page Remove page | Previous | 1 of 66 | Next |
|
small (250x250 max)
medium (500x500 max)
large ( > 500x500)
Full Resolution
All (PDF)
|
This page
All
Subset |
THE USE OF DYNAMIC SYNAPSE NEURAL NETWORKS
FOR SPEECH PROCESSING TASKS
by
Sageev Thycodam George
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BIOMEDICAL ENGINEERING)
August 2007
Copyright 2007 Sageev Thycodam George
Object Description
| Title | The use of dynamic synapse neural networks for speech processing tasks |
| Author | George, Sageev Thycodam |
| Author email | sageev@gmail.com |
| Degree | Doctor of Philosophy |
| Document type | Dissertation |
| Degree program | Biomedical Engineering |
| School | Viterbi School of Engineering |
| Date defended/completed | 2007-04-23 |
| Date submitted | 2007 |
| Restricted until | Unrestricted |
| Date published | 2007-07-26 |
| Advisor (committee chair) | Berger, Theodore |
| Advisor (committee member) |
Baudry, Michel D'Argenio, David Yamada, Walter |
| Abstract | Automated signal processing systems have become increasingly popular in the modern world. One challenge to developing such systems is the user's inability or lack of desire to provide large sample sets for model development or classification tasks. In such situations, a complete statistical description of signals to be classified is not possible. Nevertheless, life on earth (in particular neural systems) seems capable of temporal signal classification even when given only a small number of sample signals. -- In this thesis, I explore this seemingly incongruent situation.; I have developed automated signal processing systems using a relatively new advancement in neural computation, the dynamic synapse neural network (DSNN), applied to small numbers of sample signals. I have developed a speech recognition system and a speaker verification system within the well-established, and highly statistics based, area of speech processing, in order to better measure and understand how the resulting DSNN based systems generalize. -- Dynamic Synapse Neural Networks, based on biological principles, can be used to perform speech recognition and speaker verification tasks.; A single DSNN is used for the speaker verification system. Speech samples undergo a wavelet decomposition and adaptive thresholding process which generates seven pulse trains. The trains are used as input to a seven input neuron-two output neuron DSNN with full feedback to the presynapses of the input layer. Ten DSNNs with an identical configuration are used for the speech recognition system. -- For the speaker verification system, 5, 10, 15, 20 or 25 samples per pass-word were used for training. Systems were trained on the closed-set of one target speaker and seven imposters, and tested on the open-set of eight novel imposters.; For the speech recognition system, 5, 10, 15, 20 or 25 samples per pass-word were used for training. Systems were trained on one speaker and tested on fifteen speakers. -- The DSNN based speech recognition and speaker verification systems prove capable of being trained small number of sample signals. They also prove capable of generalization, with the speech recognition system classifying speech uttered by novel speakers, and the speaker verification system rejecting imposters from the open set of speakers. The results establish the validity of this approach for these speech processing tasks. |
| Keyword | neural network; speech recognition; speaker verification; dynamic synapse |
| Language | English |
| Part of collection | University of Southern California dissertations and theses |
| Publisher (of the original version) | University of Southern California |
| Place of publication (of the original version) | Los Angeles, California |
| Publisher (of the digital version) | University of Southern California. Libraries |
| Type | texts |
| Legacy record ID | usctheses-m679 |
| Rights | George, Sageev Thycodam |
| Repository name | Libraries, University of Southern California |
| Repository address | Los Angeles, California |
| Repository email | http://www.usc.edu/isd/libraries/services/ask_a_librarian/email/ |
| Filename | etd-George-20070726 |
| Archival file | uscthesesreloadpub_Volume26/etd-George-20070726.pdf |
Description
| Title | Page 1 |
| Full text | THE USE OF DYNAMIC SYNAPSE NEURAL NETWORKS FOR SPEECH PROCESSING TASKS by Sageev Thycodam George A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (BIOMEDICAL ENGINEERING) August 2007 Copyright 2007 Sageev Thycodam George |
Comments
Post a Comment for Page 1
