Browse Prior Art Database

Speech Recognition Method

IP.com Disclosure Number: IPCOM000101309D
Original Publication Date: 1990-Jul-01
Included in the Prior Art Database: 2005-Mar-16
Document File: 2 page(s) / 57K

Publishing Venue

IBM

Related People

Nishimura, M: AUTHOR

Abstract

This article describes a process of speech recognition that can simultaneously cope with isolated words from a large vocabulary and connected words from a small vocabulary, without increasing the computation time.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 73% of the total text.

Speech Recognition Method

       This article describes a process of speech recognition
that can simultaneously cope with isolated words from a large
vocabulary and connected words from a small vocabulary, without
increasing the computation time.

      The recognition system consists of an isolated-word recognizer
and a connected-word recognizer. The isolated-word recognizer has two
functions: word preselection and detail matching for the preselected
words. A one-state Hidden Markov Model (HMM) is used for the word
preselection, and a fenonic Markov model is used for the detail
matching. Since the word preselection needs to detect the ends of
utterances, the detail matching cannot begin before the end of an
utterance. The connected-word recognizer uses the same fenonic Markov
model as the isolated-word recognizer.

      The figure shows the process of this speech recognition system:
1.  Word boundaries are detected by a 'double threshold method.'
2.  Connected-word recognition is done frame-wise by means of a
'onepass algorithm.'
3.  At the same time, word preselection is done frame-wise for a
large vocabulary.
4.  When low power is detected, the word preselection results and the
connected-word recognition results are saved.
5.  If the low power continues for several frames, the word preselec
tion and the connected-word recognition process are terminated.
Isolated-word recognition is then done for the preselected words by
means of the 'Viterbi algorithm....