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

Iteratove Prototype Estimation in Vector Quantising Speech Recognition Systems

IP.com Disclosure Number: IPCOM000105951D
Original Publication Date: 1993-Sep-01
Included in the Prior Art Database: 2005-Mar-20
Document File: 2 page(s) / 46K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+2]

Abstract

[1-3] describe vector quantisation algorithms which have proven very successful for speech recognition. The prototypes are created by clustering a sample of acoustic parameter vectors. These prototypes are then used to label test data via the method of maximum likelihood. However, using held-out training data where the correct labels are known, it can be demonstrated experimentally that many frames are mislabelled by the methods of [1-3]. The following method improves the labeller by re-processing the labelling errors made on the vector-quantisation training data. The resulting labeller is more accurate on both the training data and on held-out data.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 55% of the total text.

Iteratove Prototype Estimation in Vector Quantising Speech Recognition Systems

      [1-3] describe vector quantisation algorithms which have proven
very successful for speech recognition.  The prototypes are created
by clustering a sample of acoustic parameter vectors.  These
prototypes are then used to label test data via the method of maximum
likelihood.  However, using held-out training data where the correct
labels are known, it can be demonstrated experimentally that many
frames are mislabelled by the methods of [1-3].  The following method
improves the labeller by re-processing the labelling errors made on
the vector-quantisation training data.  The resulting labeller is
more accurate on both the training data and on held-out data.

1.  Create Euclidean centroids and Gaussian prototypes via the
    clustering algorithms of [1-3].

2.  Label the vector-quantisation training date [1,2], as if it were
    test data.

3.  Extract all the parameter vectors which were mislabelled in Step
    2.

4.  Perform Euclidean clustering on the errorful data extracted in
    Step 3, using the algorithm of [1].  Do not perform the Gaussian
    clustering also described in [1].

5.  Append the Euclidean centroids obtained in Step 4 to those
    previously derived, thus obtaining an augmented set of centroids.
    This augmented set contains centroids deliberately positioned in
    the midst of those regions in the acoustic parameter space where
   ...