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

Approximate Conditional Output Probability Estimator for Speech Recognition Systems Using Discrete Parameter Acoustic Markov Models

IP.com Disclosure Number: IPCOM000099917D
Original Publication Date: 1990-Mar-01
Included in the Prior Art Database: 2005-Mar-15
Document File: 2 page(s) / 52K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+2]

Abstract

In discrete-parameter acoustic Markov source models, it is assumed that at time t, some arc at produces an observed label ft . For simplicity, it is usually assumed that (Image Omitted) which asserts that knowing the preceding arcs at-1 provides no additional information about ft once the current arc at is known, and that knowing the preceding labels ft-1 provides no additional information about ft once at is known. Although the first assertion appears reasonable, the second assertion is patently false. A much milder, and more reasonable assumption is which asserts that once the current arc at and the previous label ft-1 are known, all other preceding arcs and labels provide no further information about ft .

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 90% of the total text.

Approximate Conditional Output Probability Estimator for Speech Recognition Systems Using Discrete Parameter Acoustic Markov Models

       In discrete-parameter acoustic Markov source models, it
is assumed that at time t, some arc at produces an observed label ft
.  For simplicity, it is usually assumed that

                            (Image Omitted)

 which asserts that knowing
the preceding arcs at-1 provides no additional information about ft
once the current arc at is known, and that knowing the preceding
labels ft-1 provides no additional information about ft once at is
known.  Although the first assertion appears reasonable, the second
assertion is patently false.  A much milder, and more reasonable
assumption is which asserts that once the current arc at and the
previous label ft-1 are known, all other preceding arcs and labels
provide no further information about ft .  Although (2) is more
accurate than (1), it involves many more parameters which are
difficult to estimate reliably, and expensive to store.  This renders
(2) impractical, and (1) is usually used instead.  In this article we
describe a practical approximation of Pr(ft   at,ft-1) which is more
accurate than (1). By Baye's rule, For simplicity, assume that which
asserts that knowing ft-1 provides no additional information about
the current arc a t once the current label ft is known.  Substituting
(4) into (3) leads to By Baye's rule, Substituting (6) into (5) give...