Phoneme recognition with neural network preprocessing filters and backpropagation.
Item
-
Title
-
Phoneme recognition with neural network preprocessing filters and backpropagation.
-
Identifier
-
AAI9325063
-
identifier
-
9325063
-
Creator
-
Alankar, Sudhir.
-
Contributor
-
Adviser: John Antrobus
-
Date
-
1993
-
Language
-
English
-
Publisher
-
City University of New York.
-
Subject
-
Computer Science | Health Sciences, Speech Pathology
-
Abstract
-
Hearnet is an artificial neural network designed to simulate speaker independent speech recognition. The design of the model exploits neurophysiological findings on the afferent pathway from the outer ear to the auditory cortex. HearNet consists of a feedforward network, FilterNet, and a recurrent backpropagation network, LearnNet. FilterNet transforms the speech signal into phonemic information through cascaded, spatially distributed, neural network filters. The output of the FilterNet is fed into LearnNet which learns the successive invariances in the speech signal inorder to recognize phonemes.;The purpose of this research was to investigate the advantage of a speech recognition model based on the architecture of the human auditory system. FilterNet filters are modelled after the neurons in the cochlear nucleus and other neural bodies, that are presumed to extract salient features of speech. Preprocessing of the speech signal by FilterNet simplifies the computational load of LearnNet. Each of the parallel filters of FilterNet passes relevant acoustic feature information of speech, in compacted form, to LearnNet. The filters of the HearNet model distinguish it from known speech recognition systems which require an enormous amount of computing power and are confounded by speaker variability.;HearNet recognizes a small number of speaker independent phonemes of monosyllabic words, in isolation with a recognition rate of 100%.
-
Type
-
dissertation
-
Source
-
PQT Legacy CUNY.xlsx
-
degree
-
Ph.D.
