Browse Prior Art Database

Simplifying Speech Recognition with Individual Usage Patterns

IP.com Disclosure Number: IPCOM000117076D
Original Publication Date: 1995-Dec-01
Included in the Prior Art Database: 2005-Mar-31
Document File: 2 page(s) / 40K

Publishing Venue

IBM

Related People

Cohen, PS: AUTHOR [+3]

Abstract

Disclosed is a method for simplification of the voice recognition process through a knowledge of patterns exhibited by an individual in using letters, numbers, and words. This method is used to simplify the branching conditions which must be considered during the speech recognition process, and to identify and correct errors during the process. Examples of data repeatedly used by a particular individual, for which usage patterns may be detected, are telephone numbers, e-mail (electronic mail) IDs, account numbers, serial numbers, printer IDs, names, and social security numbers.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 87% of the total text.

Simplifying Speech Recognition with Individual Usage Patterns

      Disclosed is a method for simplification of the voice
recognition process through a knowledge of patterns exhibited by an
individual in using letters, numbers, and words.  This method is used
to simplify the branching conditions which must be considered during
the speech recognition process, and to identify and correct errors
during the process.  Examples of data repeatedly used by a particular
individual, for which usage patterns may be detected, are telephone
numbers, e-mail (electronic mail) IDs, account numbers, serial
numbers, printer IDs, names, and social security numbers.

      Preferably, the speech recognition system combines both usage
and acoustic probabilities to select the highest probability
candidate for a word or number, or to identify an error.  Since in
some instances, it may be impractical to keep a complete list of all
the phone numbers and other numeric combinations used by an
individual, Lempel-Ziv tables, n-grams, or check sums may alternately
be combined with the acoustic probabilities.  While n-grams, which
are typically bi-grams or tri-grams, usually work poorly for the
recognition of numbers from large samples, such as a telephone book,
they can be effective when used with smaller samples of numbers.
Also, they are useful in providing the spelling of names, cities,
etc.  While Lempel-Ziv techniques, which involve the use of tables
and trees, are not as effective as...