Browse Prior Art Database

Method for Compressing a Fast Match Table

IP.com Disclosure Number: IPCOM000120889D
Original Publication Date: 1991-Jun-01
Included in the Prior Art Database: 2005-Apr-02
Document File: 3 page(s) / 70K

Publishing Venue

IBM

Related People

Nishimura, M: AUTHOR

Abstract

This article describes a technique that saves memory space for the polling fast match [*], which is used for word preselection in a speech recognition system on a speech processing card (SPC).

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 78% of the total text.

Method for Compressing a Fast Match Table

      This article describes a technique that saves memory
space for the polling fast match [*], which is used for word
preselection in a speech recognition system on a speech processing
card (SPC).

      Let L1, L2, L3, ...  Ln be a label sequence produced by a
vector quantizer in response to an utterance of some unknown word.
The vector quantizer is designed on the basis of binary tree coding
(BTC).  In the polling fast match, the score Sw is estimated for each
word w by the following polling function: where Pr(L@w) is a fast
match table.  The size of v is proportional to the codebook size as
well as the vocabulary size.  Since the table requires a large amount
of memory, it is kept in the PC's memory.  It is compressed in
proportion to the active vocabulary size and the compressed table is
then transferred from the PC to the SPC.  This process is shown in
Fig. 1.

      Since the label is coded by BTC, the label number corresponds
to the structure of the codebook, as shown in Fig. 2. Let L(Bn, Bn-1
,..., B2, B1) be a label number, where Bi is a bit representing a
branch at the i-th stage in the codebook. The table Pr(L@w) is
compressed by the following equation: where i=log2 C, and
C(=2,4,8,...) is a compression ratio that is determined by the
relation between the available memory on the SPC and the size of the
fast match table. This procedure is illustrated in Fig.  3.  On the
other hand, bits Bi,...,B1 are...