Browse Prior Art Database

Addword Through Correction of Recognition Errors

IP.com Disclosure Number: IPCOM000108537D
Original Publication Date: 1992-Jun-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 2 page(s) / 81K

Publishing Venue

IBM

Related People

De Gennaro, SV: AUTHOR [+4]

Abstract

Disclosed is a method for adding new words to a speech recognition system in a way that is natural and efficient for the user - through the correction of recognition errors during dictation.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Addword Through Correction of Recognition Errors

       Disclosed is a method for adding new words to a speech
recognition system in a way that is  natural and efficient for the
user - through the correction of recognition errors during dictation.

      Speech recognition systems generally have finite vocabularies,
requiring some customization for each user, typically involving the
construction of acoustic models (based  on inputs such as the
spelling, sounds-like spelling and/or one or more sample utterances
of the word), and the assignment of language model probabilities
reflecting the likely linguistic context of the word's usage.  By
correcting an error during the actual use of the recognizer, the
required acoustic and linguistic inputs can be captured.

      In correction during use, the user does not need to predefine
additional words to the system.  When the system makes an error, due
to an out-of-vocabulary word or the misrecognition of an
in-vocabulary word, the user simply corrects the error.

      The decision to add an acoustic and/or linguistic
representation is then handled automatically by the speech
recognition system, without further intervention of the user.  This
is a natural, simple method for the user, with vocabulary
augmentation handled as a side-effect of error correction, which is
required in any case for  a correct document.

      The scheme can be implemented by:
1.   The user corrects an error, providing a correct spelling for the
word or phrase, and optionally a sounds-like spelling (i.e.,
"triple-A" for "AAA") for a given acoustic and textual segment.
2.   If the error is in-vocabulary (i.e., a normal recognition
error), the system can then decide to:
      o    Ignore  the error from the point of adding a word to the
user's vocabulary.
      o    Use an adaptation scheme to adjust existing acoustic and
linguistic models.
      o    Add a new, parallel acoustic and linguistic model to...