Previous: SOM-PAK:
Up: NEURAL NETWORK SIMULATION TOOLS
Next: Valentino
Previous Page: SOM-PAK:
Next Page: Valentino
General Information
- Name
- TDNN (Time Delay Neural Network)
- Source
- i.n.a. (= ``information not available'')
- Price
- free of charge
- Target usage
- demonstration example; beginners
Model
- Network Architecture
-
The network has one fixed, predefined topology for the recognition of
the ten spoken digits 0-9 from continuous speech; the input layer
consists of 16 x 11 units, i. e. the outputs of 15 frequency-domain
bandpass-filters and one zero-crossing detector sampled at 10 equally
spaced time points and one word endpoint.
- Paradigms
- Time Delay Neural Network
- Limitations
-
Training takes exactly 64 iterations; training method not described.
(see also network description above)
Implementation
- Display
- monochrome
- User Interface
- shell
- Execution
- By logging of commands
- Data Format
-
ASCII files containing the paths and names of speech data files;
10 speech data files containing digitized speech data for an example of
the spoken digits.
- Evaluation
-
Scoring programs that calculate recognition accuracy based on an area
and on a threshold criterion are enclosed with the package.
- Performance
-
System Requirements
- Hardware
- PC
- Operating System
- DOS
- Display
- monochrome
- Language
- C, FORTRAN
Installation and Documentation
- Installation
- By lha in DOS
- Sources available
- Programs written and poorly commented in FORTRAN
and C are included in the package.
- Implementation language
- C, FORTRAN
- Documentation
-
Only a very short description of the files and programs is available.
The simulation programs are broken up into five sub-directories.
Front contains the feature extraction program. Prelearn
contains the time delaying program. Learn contains the network
training program. Recognize contains the testing program.
And Useful contains several scoring programs, a segmentation
program to isolate single digits from continuous speech based on
network recognition,
and a program to re-align (rejustify) the endpoints of the training
digits.
- Support
- none
- User Groups
-
Previous: SOM-PAK:
Up: NEURAL NETWORK SIMULATION TOOLS
Next: Valentino
Previous Page: SOM-PAK:
Next Page: Valentino