Browse Prior Art Database

Estimating the Acoustic Scores in a Speech Recognition System Using Context Dependent Acoustic Markov Models

IP.com Disclosure Number: IPCOM000109879D
Original Publication Date: 1992-Sep-01
Included in the Prior Art Database: 2005-Mar-24
Document File: 1 page(s) / 45K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+4]

Abstract

In one prominent approach to speech recognition, context-dependent acoustic Markov models are used. Corresponding to each arc in the inventory of arcs is a collection of context-dependent prototypes: the appropriate prototype is selected according to the context of the arc. The mapping from context to prototype is not always optimal - especially if it is obtained empirically from training data. Inaccuracies in the mapping and/or the prototypes can lead to a poor score being ascribed to a frame of speech, when a good score would be obtained if a different prototype of the same arc were used instead. The invention below takes advantage of this observation to smooth the acoustic scores, thereby reducing the recognition error rate.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 81% of the total text.

Estimating the Acoustic Scores in a Speech Recognition System Using Context Dependent Acoustic Markov Models

       In one prominent approach to speech recognition,
context-dependent acoustic Markov models are used.  Corresponding to
each arc in the inventory of arcs is a collection of
context-dependent prototypes: the appropriate prototype is selected
according to the context of the arc.  The mapping from context to
prototype is not always optimal - especially if it is obtained
empirically from training data.  Inaccuracies in the mapping and/or
the prototypes can lead to a poor score being ascribed to a frame of
speech, when a good score would be obtained if a different prototype
of the same arc were used instead.  The invention below takes
advantage of this observation to smooth the acoustic scores, thereby
reducing the recognition error rate.

      The acoustic score associated with arc A in  context C for
vector X is computed as follows.
(1)  Using the context-to-prototype mapping, determine the prototype
P which represents arc A in context C.
(2)  Compute the acoustic score of vector X when modelled by the
prototype P of Step 1.  Denote this score as S.
(3)  Perform Step 4 for each prototype Q associated with arc A.
(4)  Compute the acoustic score of vector X when modelled by
prototype Q.
(5)  Locate the maximum score obtained in Step 5 and denote it as M.
(6)  Compute a smoothed score for vector X and arc A in context C as
W.S + (1 - W).M

    ...