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

Adaptation of Acoustic Prototypes by Partial Tying

IP.com Disclosure Number: IPCOM000109694D
Original Publication Date: 1992-Sep-01
Included in the Prior Art Database: 2005-Mar-24
Document File: 3 page(s) / 95K

Publishing Venue

IBM

Related People

Bakis, R: AUTHOR [+4]

Abstract

Prototypical spectral vectors based on plentiful reference speech are perturbed using new but sparse speech data. The idea is to fit a parsimonious parametrization of the change in the spectrum.

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

Adaptation of Acoustic Prototypes by Partial Tying

       Prototypical spectral vectors based on plentiful
reference speech are perturbed using new but sparse speech data.  The
idea is to fit a parsimonious parametrization of the change in the
spectrum.

      A 'Z-label' prototype for a fixed phone in d-dimensional space
(e.g., 50-dimensional) is a mixture of k (e.g., 10) Gaussian
densities.  We adapt such a prototype by
o    Retaining the k subclass weights p1,...,pk and retaining the kd
Gaussian variances         .
o    Moving the prototype mean (= weighted average of subclass means)
in the direction of the average of the new data vectors that belong
to the prototype.
o    Moving each subclass mean (= mean of a Gaussian component of the
mix) in the direction of the corresponding subclass average based on
new data.
o    Constructing the new subclass means from the above quantities
combined with a damped version of the change between the old
deviations of subclass means from the prototype mean and the
corresponding deviations based on new data.

      Depending on the damping factor, a varying amount of 'tying' is
possible between the adaptive movements of the different subclass
means.

      Viterbi align new data so as to identify the phone and the
subclass of each new spectral vector.  For each phone proceed as
follows.
o    Denote the prototype mean by m, the subclass means by mi so that

                            (Image Omitted)

      Denote the mean of the new data by X and the subclass means of
the new data by Xi for i=1,...,k.  Compute an intermediate estimate
of the new prototype mean as
      where 0 < g < 1.
o    For each subclass i compute an intermediate estimate of the new
subclass mean as
      where 0 < r < 1.
o   For each subclass i compute the subclass deviation vector
    and the corresponding intermediate deviation vector and let
     be the change in the deviation vectors.
o   Combine these quantities to form the final subclass estimates as
     where 0 < e < 1.  The ca...