Browse Prior Art Database

Isolated Word-Recognition Method

IP.com Disclosure Number: IPCOM000061766D
Original Publication Date: 1986-Sep-01
Included in the Prior Art Database: 2005-Mar-09

Publishing Venue

IBM

Related People

Authors:
Nishimura, M Okochi, M [+details]

Abstract

This article describes adoption of probabilistic definition of word boundaries and a matching window in a new isolated word-recognition technique using a Hidden Markov Model (HMM) with duration distribution (Duration Model). The Duration Model uses the duration of a word in an explicit form compared with the conventional HMM. In our proposal, the following improvements are introduced into the Duration Model. (1) To reduce the recognition errors caused by the erroneous word boundaries, the word boundaries are defined by using the probabilities which are obtained as the recognition results for silence followed by a word onset. (We call the probabilistic model "Silence Machine".) (2) The matching window is constrained by using the duration distributions which are obtained by training.