Browse Prior Art Database

Improvement in Noise Rejection by Limiting Loops in Markov Models

IP.com Disclosure Number: IPCOM000112619D
Original Publication Date: 1994-Jun-01
Included in the Prior Art Database: 2005-Mar-27
Document File: 2 page(s) / 64K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+5]

Abstract

An algorithm is disclosed to improve the noise rejection in Markov models by limiting the number of times a self-loop transition is taken in the Viterbi algorithm.

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

Improvement in Noise Rejection by Limiting Loops in Markov Models

An algorithm is disclosed to improve the noise rejection in Markov
models by limiting the number of times a self-loop transition is
taken in the Viterbi algorithm.

      In speech recognition systems, a commonly-used type of Markov
model has the structure shown in the Figure below [*].

    /\     /\     /\     /\     /\
      /  \   /  \   /  \   /  \   /  \
      \  /   \  /   \  /   \  /   \  /
       \/     \/     \/     \/     \/
        0------0------0------0------0------0

      Each state has a self-loop which accommodates the variation in
the length of different utterances of the same word.  The use of such
self-loops in the model can be detrimental if the recognition system
is required to work in a noisy environment and rejects inputs that
are a result of noise.  This is sometimes referred to as the "mumble
rejection" problem.

      An input corresponding to noise may trigger the recognition of
a word if the noise results in a long sequence of acoustic labels
that have a high probability of being produced by one of the
self-loops in the model.  The model assigns a high probability to the
noise input by looping through the same self-loop over and over
again.

      The invention relates to a modification of the Viterbi
algorithm.  This modification has the effect of limiting the number
of times a self-loop can be traversed, thereby reducing this spurious
recognition problem.

      Let
 s sub 1 , s sub 2 , ...  , s sub n denote the states of the model.
Let
 x sub 1 , x sub 2 , ...  , x sub m denote the label sequence.  Let
p(i,j) denote the probability assigned to label x sub j by a
transition emanating from state s sub i.  Let v(i,j) denote the
Viterbi score associated with state index i and label index (or frame
index) j.

      In the normal Viterbi calcu...