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

Using Conditional Null-Arc Probabilities in Markov Word Models to Prevent Deletion of Phones

IP.com Disclosure Number: IPCOM000102674D
Original Publication Date: 1990-Dec-01
Included in the Prior Art Database: 2005-Mar-17
Document File: 2 page(s) / 55K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+4]

Abstract

In speech recognition systems using discrete-output acoustic Markov models, the transition and output probabilities are usually independent of the past and depend only on the current state and arc. In this environment, some undesirable consequences can occur. The null arcs, which allow portions of the model to be deleted, are intended to allow for shortening of a phoneme, but actually allow shortening to the point of deletion. This results in errors. This invention presents a remedy to the deletion problem.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 65% of the total text.

Using Conditional Null-Arc Probabilities in Markov Word Models to Prevent Deletion of Phones

       In speech recognition systems using discrete-output
acoustic Markov models, the transition and output probabilities are
usually independent of the past and depend only on the current state
and arc.  In this environment, some undesirable consequences can
occur.  The null arcs, which allow portions of the model to be
deleted, are intended to allow for shortening of a phoneme, but
actually allow shortening to the point of deletion.  This results in
errors.  This invention presents a remedy to the deletion problem.

      It is desirable to penalize null-arc transitions heavily when
they are being used to delete an entire phoneme, but to leave them
unchanged when they are being used "properly" to shorten a phoneme.
Deletion of a phoneme is indicated when the nearby labels are
atypical of the arcs being bypassed, whereas shortening is indicated
when the opposite holds.

      The probability of a null-arc transition is modified as
follows.  Let Q denote the probability of a null-arc transition from
a given node N, as computed during training. The modified probability
Q' at time t is defined as where Pr(Ft-1 ¯ N) denotes the probability
of the preceding label Ft-1 being produced when a non-null arc is
taken from node N.  If there are no non-null arcs originating at node
N, the ratio in Equation (1) is undefined and Q' is defined instead
as Q' =Q.

      The...