Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Classifying Hypernasality for Children with a Cleft Palate using a Convolutional Neural Network

Svensson, Rebecca LU (2022) In Master's Theses in Mathematical Sciences FMSM01 20221
Mathematical Statistics
Abstract
In Sweden, about 150-200 children annually are born with some form of
cleft lip and/or palate, making it the most common facial malformation in
the country. Treatment often involves one or more surgeries and the speech
development is followed up by a speech pathologist from the first year of life.
One of the most common speech deviation for children with a cleft palate is
hypernasality, which comes from the condition velopharyngeal insufficiency
(VPI), which means that the soft palate cannot close the passage between
the throat and nasal cavities properly. To classify the severity of VPI and
hypernasality is hard, since listeners have different internal standards. This
makes it important to develop an independent method to... (More)
In Sweden, about 150-200 children annually are born with some form of
cleft lip and/or palate, making it the most common facial malformation in
the country. Treatment often involves one or more surgeries and the speech
development is followed up by a speech pathologist from the first year of life.
One of the most common speech deviation for children with a cleft palate is
hypernasality, which comes from the condition velopharyngeal insufficiency
(VPI), which means that the soft palate cannot close the passage between
the throat and nasal cavities properly. To classify the severity of VPI and
hypernasality is hard, since listeners have different internal standards. This
makes it important to develop an independent method to classify the severity.

This thesis studied two existing deep learning methods from another master’s
thesis on new data to see if it would be able to classify the severity of VPI,
which is classified on a three point scale. The methods did not work well
on the new data, but could be improved a bit by better data processing and
some changes in the methods. The best performance for classifying VPI was a
VGGish network which gave a file-wise accuracy of 57.11 %. A method to try
and classify hypernasality was investigated as well. The best method found
in order to classify hypernasality, which is classified on a four point scale, was
to look at the vowels and to use mel spectrograms in a Convolutional Neural
Network (CNN). Nasometry data was also given and together with the mel
spectrograms an accuracy of 52.38 % was reached. (Less)
Please use this url to cite or link to this publication:
author
Svensson, Rebecca LU
supervisor
organization
course
FMSM01 20221
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Cleft Lip, Cleft Palate, Speech, CNN, VGGish, Deep Learning
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMS-3456-2022
ISSN
1404-6342
other publication id
2022:E60
language
English
id
9093969
date added to LUP
2022-06-29 14:41:03
date last changed
2022-07-20 13:28:58
@misc{9093969,
  abstract     = {{In Sweden, about 150-200 children annually are born with some form of
cleft lip and/or palate, making it the most common facial malformation in
the country. Treatment often involves one or more surgeries and the speech
development is followed up by a speech pathologist from the first year of life.
One of the most common speech deviation for children with a cleft palate is
hypernasality, which comes from the condition velopharyngeal insufficiency
(VPI), which means that the soft palate cannot close the passage between
the throat and nasal cavities properly. To classify the severity of VPI and
hypernasality is hard, since listeners have different internal standards. This
makes it important to develop an independent method to classify the severity.

This thesis studied two existing deep learning methods from another master’s
thesis on new data to see if it would be able to classify the severity of VPI,
which is classified on a three point scale. The methods did not work well
on the new data, but could be improved a bit by better data processing and
some changes in the methods. The best performance for classifying VPI was a
VGGish network which gave a file-wise accuracy of 57.11 %. A method to try
and classify hypernasality was investigated as well. The best method found
in order to classify hypernasality, which is classified on a four point scale, was
to look at the vowels and to use mel spectrograms in a Convolutional Neural
Network (CNN). Nasometry data was also given and together with the mel
spectrograms an accuracy of 52.38 % was reached.}},
  author       = {{Svensson, Rebecca}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master's Theses in Mathematical Sciences}},
  title        = {{Classifying Hypernasality for Children with a Cleft Palate using a Convolutional Neural Network}},
  year         = {{2022}},
}