Speech Processing
by Computer
LECTURE 6b
NEURAL NETWORKS
Objectives
By the end of the
session you should:
q
be able to describe three applications of neural networks in speech and
language
q
have a sense of how neural network models may be used to process
language data, and how they may be trained and tested
q
be able to describe some strengths and weaknesses of the three
applications, particularly with reference to symbolic alternatives
Outline
1. Learning
past-tense morphology
q
Generating the phonological representation of the past-tense form of
given verb roots
q
Learned from examples
q
Tested for generalisation ability
2. Lexical access
q
Identifying words from phonetic transcription
q
Lexical effect, lexical segmentation, time course of lexical access
3. Phoneme
recognition
q
Identifying phoneme probabilities from acoustic input
q
Learned from large annotated speech corpus
4. Issues
q
Generalisation abilities of networks
q
Predictive power of networks as cognitive model
q
Practical model of some cognitive processes
Reading
Recommended:
q
“Connectionism”, Stanford Encyclopedia of Philosophy, http://plato.stanford.edu/entries/connectionism/
Background:
D.E.Rumelhart & J.L.McClelland, On learning the past tenses of English verbs, in Parallel Distributed Processing, Vol 2, Chapter 18, MIT Press 1986, pp216-271.
J.L. McClelland & J.L. Elman , Interactive processes in speech perception: the TRACE model, in Parallel Distributed Processing, Vol 2, Chapter 15, MIT Press 1986, pp58-121.
A.J. Robinson, An application of recurrent nets to phone probability estimation, IEEE Transactions on neural networks, 5 (1994), pp298-305.