DPREC

NAME

dprec - whole word template recogniser

SYNOPSIS

dprec (-l|-r|-p) (-d database) (-f num) (-i item) (-t token) (-a) (-c) (-s) (-S top) (-n nearest) (-F features) file(s)

DESCRIPTION

dprec is an isolated word recognizer implemented using a dynamic programming algorithm. It maintains a set of templates (the database) from one run to the next, so that it can add templates when it fails to recognise a word correctly. If a template causes many mis recognitions then it can be forgotten from the database.

The dp program operates using a Euclidean metric between vectors of a "Coefficient" (CO) data set in the files (identified by the "-i item" switch). To identify a token, the 'token' field in the main header of the input file is used.

Options and their meanings are:

-I Identify program name and version number.

-l Learn mode. If an input word is not recognised correctly, then add it to the database of templates.

-r Recognition mode. Find the closest token to each input file using the database of templates. Default.

-p Purge database of unused templates. (Non functional)

-d database Specify the file to use to collect templates. This is just an SFS file. Default: dprec.sfs.

-f num Cause templates that cause more than num misrecognitions to be forgotten.

-i item Select input item number.

-a Save all templates seen during training.

-c Clear all scores of recognition success/failure (see -f switch)

-s Print scores of recognition success/failure for each template.

-S topnum Display list of topnum best matches and distances.

-n num Use a k nearest neighbour voting system with num neighbours. Max number of neighbours=100.

-F featurefile Perform voting by accumulating scores by features rather than by tokens. 'featurefile' contains a feature breakdown for each template token name. Each line contains the token name and list of features separated by spaces, e.g. "ba BILABIAL VOICED STOP".

INPUT ITEMS

CO.xx Any Coefficient item.

VERSION/AUTHOR

1.3 Mark Huckvale

BUGS

Although the database of templates looks like an SFS file, in fact the header token field is used to maintain the template match history, so items may not be added or deleted by external programs.
Fri Jul 09 14:54:38 2004