Browse Prior Art Database

Method for Creating a Label-Smoothing Matrix by using Fine-Pitch Data

IP.com Disclosure Number: IPCOM000114469D
Original Publication Date: 1994-Dec-01
Included in the Prior Art Database: 2005-Mar-28
Document File: 2 page(s) / 41K

Publishing Venue

IBM

Related People

Sugawara, K: AUTHOR

Abstract

A technique is described whereby the label-smoothing matrix used in the N-segment label histogram method is created from a small quantity of training speech data. To overcome the sparseness of the data, this method uses fine-pitch data. The frame-shift (pitch) of the analyzing window is 1.6 ms whereas that of a normal system is 12.8ms. The label-smoothing value is obtained by calculating co-occurrence data weighted by the relative positions of the frames. This method does not require multiple utterances for the same set of training words (1,2).

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 67% of the total text.

Method for Creating a Label-Smoothing Matrix by using Fine-Pitch Data

      A technique is described whereby the label-smoothing matrix
used in the N-segment label histogram method is created from a small
quantity of training speech data.  To overcome the sparseness of the
data, this method uses fine-pitch data.  The frame-shift (pitch) of
the analyzing window is 1.6 ms whereas that of a normal system is
12.8ms.  The label-smoothing value is obtained by calculating
co-occurrence data weighted by the relative positions of the frames.
This method does not require multiple utterances for the same set of
training words (1,2).

The procedure is as follows:
  1.  Get the features of the data, using
      o  10 kHz sampling PCM
      o  An analyzing window of 25.6 ms with 1.6 ms frame-shift
      o  DFT and 19-channel band-pass filtering
  2.  Label the data by using a predefined codebook (codebook size
64).
  3.  Calculate the co-occurrence.  If labels 'i' and 'j' appear
within
       k%(0 le k le 3) frames, the co-occurrence count of 'i' and
'j',
       c(i, j), is increased by 2 sup <%(3-k)> .  This means that the
       nearer the label positions are, the higher the co-occurrence
       score is.
  4.  Iterate steps 1-3 for all training data.  Normalize the
       co-occurrence count so that it becomes a probability matrix.
To
       avoid zero-probability problem, entries below some small
       thre...