Biologically Plausible Neural Network Models of Prelexical Speech Processing
Keywords
Loading...
Authors
Issue Date
2021-06-18
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Speech recognition requires mapping highly variable sound waves into concrete units of speech. By
means of this mapping, the correct words can be reconstructed. However, which exact speech units are
learned for prelexical processing has been an area of debate. To investigate this matter, we simulate a
low-level area in the auditory pathway of the brain, by means of a spiking neural network (SNN), that
receives its input from the cochlea. This SNN learns through the unsupervised spike-timing-plasticity
learning paradigm (STDP). It is hypothesised that the representations learned by the SNN are best
compared to phones because the network receives low-level input. The responses of SNNs to speci c
stimuli are analysed in order to investigate the representations learned by these types of networks.
Responses of SNNs are compared by employing a custom-de ned `response discrepancy' measure. This
measure computes the mean distance between the average responses, of for example di erent SNNs, to
the same class of stimuli, thus quantifying their dissimilarity. The response discrepancy measure is used
to compare the responses of a single SNN to two di erent test sets, as well as to compare the responses
of two di erently trained SNNs to the same test set. This analysis of the response dissimilarities
showed no generalisation of the network's responses to phones or words uttered by di erent speakers
from di erent dialects or genders. This suggests that the hypothesised unit of phones is not learned
by the SNN after learning.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen