Browse Prior Art Database

Estimating the Statistics in Hidden Markov Word Models When the Observations Are Correlated in Time

IP.com Disclosure Number: IPCOM000038523D
Original Publication Date: 1987-Jan-01
Included in the Prior Art Database: 2005-Jan-31
Document File: 2 page(s) / 19K

Publishing Venue

IBM

Related People

Balh, LR: AUTHOR [+4]

Abstract

In a speech recognition system using Markov models, each Markov model is characterized by transition probabilities and label output probabilities. That is, during a training session, a probability is computed for the occurrence of each transition in each Markov model and, for a label output at a given transition, a respective probability is computed. To reduce the number of probabilities in the models and thereby simplify calculations, it is assumed that the probability of a particular label output being next is dependent on model state and not on previous labels, although there is, in fact, a correlation of label outputs over time. To account for adverse effects resulting from the use of the simplifying assumption, the present invention treats each observed label output (which is a scalar) as a probabilistic vector of labels.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 51% of the total text.

Page 1 of 2

Estimating the Statistics in Hidden Markov Word Models When the Observations Are Correlated in Time

In a speech recognition system using Markov models, each Markov model is characterized by transition probabilities and label output probabilities. That is, during a training session, a probability is computed for the occurrence of each transition in each Markov model and, for a label output at a given transition, a respective probability is computed. To reduce the number of probabilities in the models and thereby simplify calculations, it is assumed that the probability of a particular label output being next is dependent on model state and not on previous labels, although there is, in fact, a correlation of label outputs over time. To account for adverse effects resulting from the use of the simplifying assumption, the present invention treats each observed label output (which is a scalar) as a probabilistic vector of labels. In speech recognition using Markov models, statistics applied in determining transition and label output probabilities are typically computed from the well-known forward-backward algorithm which attempts to find the maximum-likelihood estimates under the above assumption. In reality, speech labels are strongly correlated in time and the observed labels do, in fact, depend on the previous labels as well as the hidden state. Consequently, it cannot be assumed that the maximum-likelihood statistics are optimal for speech recognition since they have been computed under a false assumption. Experimental evidence confirms that such statistics are not optimal. In this regard, the number of consecutive identical label outputs observed are generally much greater than those predicted by the models with statistics based on the above-noted assumption. The algorithm outlined herein corrects for this discrepancy by treating each observed label output (a scalar) as a probabilistic vector of labels. Given a sequence of identical labels, this modification has the effect o...