Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Method for Constructing Alternate Word Lists

IP.com Disclosure Number: IPCOM000108349D
Original Publication Date: 1992-May-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 2 page(s) / 99K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+4]

Abstract

This invention provides a method for constructing alternate word lists in an isolated speech recognition system to make error correction easier.

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

Method for Constructing Alternate Word Lists

       This invention provides a method for constructing
alternate word lists in an isolated speech recognition system to make
error correction easier.

      The IBM speech recognizer (1) takes speech as input and
produces an ordered list of the words recognized.  Some of these
words are incorrect.  When the user indicates that a word is
incorrect, the system produces a list of alternate words, which the
user can use to make the correction.  To make error correction
easier, it is important to maximize the likelihood that the correct
word appears on the list and the likelihood that the correct word
appears near the beginning of the list.  Below are described several
methods for constructing alternate word lists that seek to maximize
these likelihoods without requiring a burdensome amount of additional
computation.

      All of these methods make use of the internal state of the
decoder.  In particular, for each sentence, the decoder builds a set
of paths that consist of a hypothesized list of the words of the
sentence being decoded.  For each word on each path, the most likely
starting time is known.  Also available for each word on each path
are the cumulative log likelihood of the path up to and including the
word, according to the language model, according to the acoustic
match, as well as a composite log likelihood.  The methods disclosed
below make use of this information.

      A second source of information being used is the decoding
history.  This is a list of all word corrections made over some prior
observation period, and contains, for each correction, the decoded
word and the correct word.

      In all cases, the information is used to generate a list of
alternate words, to assign a score to each and to order the list by
decreasing score.
Methods for Generating the List of Alternate Words
      1.   Take all the words that appear on any path whose most
likely starting time is the same as that of the incorrect word and
along which the words to the left of the decoded word are identical
to those in the decoded sentence ("filtered").
      2.   Take all the words that appear on any path and whose most
likely starting time is the same as that of the incorrect word
("unfiltered").
      3.   Re-run the decoder on the sentence containing the error,
this t...