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

Automatic Correction of Viterbi Misalignments

IP.com Disclosure Number: IPCOM000100121D
Original Publication Date: 1990-Mar-01
Included in the Prior Art Database: 2005-Mar-15
Document File: 3 page(s) / 104K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+4]

Abstract

A technique is described whereby an algorithm, used with speech recognition devices, can automatically correct Viterbi misalignment between the letters and phonemes found in the spelling and pronunciation of words.

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

Automatic Correction of Viterbi Misalignments

       A technique is described whereby an algorithm, used with
speech recognition devices, can automatically correct Viterbi
misalignment between the letters and phonemes found in the spelling
and pronunciation of words.

      Determining the phonetic repertoire of each of the letters of
the alphabet is important in studies of linguistics and phonetics, so
as to provide objective information on the many ways in which letters
can be pronounced.  The repertoire is also important in automatic
speech recognition where it plays a central role in the automatic
construction of syllabaries and in the automatic determination of the
pronunciation of words from their spellings.

      The concept described herein relies on Viterbi alignment [1] to
determine the correspondence between the letters in the spelling of a
word and the phonemes in its pronunciation.  However, it is possible
that several different alignments have identical Viterbi path
probabilities.  In these cases, the Viterbi algorithm may fail to
find the correct alignment, thereby producing an alignment which is
grossly in error.  An example could be in the word "weighed",
pronounced "W-AY-D", and aligned as follows by the Viterbi algorithm:
           w   e   i   g   h   e   d
           W   .   .   .   .   AY  D
whereas, the correct alignment is:
           w   e   i   g   h   e   d
           W   AY  .   .   .   .   D
Note that both alignments have identical Viterbi probabilities
because both involve identical letter-phoneme pairs.  However, only
the letter alignment is consistent with the alignment for "weigh":
           w   e   i   g   h
           W   AY  .   .   .

      The first alignment represents a gross error which can ruin
later attempts to construct parsimonious syllabaries and effective
spelling- to-sound rules.  For these reasons, it is necessary to
correct faulty alignments and to perform the corrections in an
automatic manner so as to preserve the semi-automatic nature of the
method.  Therefore, the concept described herein provides an
algorithmic solution to this problem.

      First, it is assumed that spelling/phoneme alignments have been
previously obtained.  Therefore, the steps of the
alignment/correction algorithm are as follows:
  Step 1. Create probabilistic spelling-to-sound rules in the form of
an idiot-system.
  Step 2. For each word W, which has a spelling/phoneme Viterbi...