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

Correction for Bias in Prototype Estimation for Vector-Quantizing Speech Recognition Systems

IP.com Disclosure Number: IPCOM000105923D
Original Publication Date: 1993-Sep-01
Included in the Prior Art Database: 2005-Mar-20
Document File: 2 page(s) / 55K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+2]

Abstract

Acoustic labels for speech recognition systems are sometimes created via the algorithms of [1,2]. These labels are determined form prototype obtained by clustering some training parameter vectors, and then estimating the means and variances of all the vectors in each cluster so obtained. The fact that the training data is being used twice, - first for clustering, and then for parameter estimation - leads to biassed estimates of the prototype parameters. In particular, the variances are underestimated.

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

Correction for Bias in Prototype Estimation for Vector-Quantizing Speech Recognition Systems

      Acoustic labels for speech recognition systems are sometimes
created via the algorithms of [1,2].  These labels are determined
form prototype obtained by clustering some training parameter
vectors, and then estimating the means and variances of all the
vectors in each cluster so obtained.  The fact that the training data
is being used twice, - first for clustering, and then for parameter
estimation - leads to biassed estimates of the prototype parameters.
In particular, the variances are underestimated.

      The procedure below reduces the bias in a manner which requires
no additional training data, nor any significant increase in
computation.

      Because the cluster centers in [1,2] are deliberately
positioned so as to minimize the average distance between the
clustered vectors and their cluster centers, the observed distances
form the clustered vectors to the final prototype centers are not
typical of what would be observed with independent vectors; the
distances are smaller.  Therefore, the prototype variances are
underestimated.  This usually means that during clustering, the
computed likelihood is too high for the nearest cluster, and too low
for all the others; the bias being greatest where the cluster sizes
are smallest.

      For this reason when small cluster sizes are involved, it is
advantageous during clustering to reduce the highest likelihood
before computing the relative cluster likel...