Browse Prior Art Database

Fast Acoustic Match Using Precomputed Lists

IP.com Disclosure Number: IPCOM000037254D
Original Publication Date: 1989-Dec-01
Included in the Prior Art Database: 2005-Jan-29
Document File: 1 page(s) / 12K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+5]

Abstract

The fast match is an important processing step in isolated as well as continuous speech recognition which determines a small subset of the vocabulary that matches well with a segment of the input signal. This is organized as a search through a tree obtained from a phonetic transcription of the words in the vocabulary (see [*] for details). We propose below a new scheme for determining a short list of acoustically similar words that match well with a segment of the input acoustics using a very fast version of the fast match scheme in [*] as an initial step. This is based on the idea that given any word one can precompute a list of acoustically similar words and use these as the short list of words to be input to the detailed match part of the decoder.

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

Page 1 of 1

Fast Acoustic Match Using Precomputed Lists

The fast match is an important processing step in isolated as well as continuous speech recognition which determines a small subset of the vocabulary that matches well with a segment of the input signal. This is organized as a search through a tree obtained from a phonetic transcription of the words in the vocabulary (see [*] for details). We propose below a new scheme for determining a short list of acoustically similar words that match well with a segment of the input acoustics using a very fast version of the fast match scheme in [*] as an initial step. This is based on the idea that given any word one can precompute a list of acoustically similar words and use these as the short list of words to be input to the detailed match part of the decoder.

The algorithm consists of two stages. First, the tree search algorithm in [*] is used to determine a short list of about 30 or less words that match well with the given acoustics. This is done with the thresholds set to be only slightly lower than the score of the best word found so far. This results in the search tree being pruned heavily and typically only about one percent of the nodes in the tree are evaluated during this stage. The correct word will be missing from this short list typically about 15 percent of the time. This is, of course, unacceptable. Now we invoke the second stage of the algorithm.

We associate a list of words with every word in the vocabulary. This list is precomputed once for all and for a given word it consists of words that are acoustically similar to that word. A method for computing such a list is described below. In the second stage of the new fast acoustic match algorithm we retrieve t...