Browse Prior Art Database

Candidate Words Proselection Using Bidirectional Prediction

IP.com Disclosure Number: IPCOM000037010D
Original Publication Date: 1989-Nov-01
Included in the Prior Art Database: 2005-Jan-29
Document File: 3 page(s) / 26K

Publishing Venue

IBM

Related People

Sugawara, K: AUTHOR

Abstract

A technique is described whereby the 'one-pass' method in continuous speech recognition is accelerated. Two methods--'beam search' and 'polling fast match'--are combined and applied to the conventional one-pass matching method to reduce the computation time.

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 60% of the total text.

Page 1 of 3

Candidate Words Proselection Using Bidirectional Prediction

A technique is described whereby the 'one-pass' method in continuous speech recognition is accelerated. Two methods--'beam search' and 'polling fast match'--are combined and applied to the conventional one-pass matching method to reduce the computation time.

The conventional one-pass matching method is as follows: Definitions: i=1,...,N : frame number of inputs k=1,...,K : reference pattern number j=1,...,J(k) : frame number of a reference pattern k where J(k) is the length of the

reference pattern k p(i,j,k) : local matching score p(i,j,k) : global matching score
l. Initialize: P(1, j, k)=p(1,1,k)*...*p(1,j,k)
2. a) For i=2,...,N do steps 2b->2e

b) For k=1,...,K do steps 2c->2e

c) P(i,1,k)=p(i,1,k)*

max(P(i-1,1,k), P(i-1, J(k'), k')

where k is connected to k')

d) For j=2,...,J(k) do step 2e

e) P(i,j,k)=p(i,j,k)*

max(P(i-1,j,k),

P(i-1,j-1,k),

P(i-1,j-2,k)) 3. Trace back using the array P(i,j,k) and get the list of words.

To eliminate unlikely paths, the beam search method is used. This method cuts relatively low-scoring grid points during computation. Whether a grid point remains in the beam and is extended further is determined by comparing the relative score M/P(i,j,k) and the beam threshold B, where: P(i,j,k) : global matching score

M : Maximum score at time i-1

By using polling fast match, the prediction of the score's growth is obtained for each grid point. The polling is done once in each polling interval R (several frames) for D consecutive frames starting at i+1. The polling scheme is as follows: h(j, 1) : similarity of word j and label 1, that is, the normalized number of occurrences of label 1 in word j computed as in an isolated-word case f(i, j) : results of polling between fr...