WawaSimi: Classification Techniques for Phonological Processes
Identification in Children from 3 to 5 Years Old
Braulio Baldeon, Renzo Ravelli and Willy Ugarte
a
Universidad Peruana de Ciencias Aplicadas (UPC), Lima, Peru
Keywords:
Phonological Processes, Language Development Level, Language Acquisition, Machine Learning.
Abstract:
In the context of the pandemic we are living in, most of the interactions between kindergarten-aged children
has decreased, meaning that their language development might be slowed down. Our work presents a machine
learning-based method for the classification of phonological processes and a corpus with a total of 3,324
audios, being 40% of them audios with a correct pronunciation, and the remaining 60% wrong. One of the
main problems encountered when trying to perform children speech recognition in Spanish, is, to the best
of our knowledge, the lack of a corpus. 329 audios were collected from 20-30 years old adults and a voice
conversion technique was applied in order to generate the required audios for the corpus construction. A
modified AlexNet was trained to classify if a word was correctly pronounced, if the audio was classified as
mispronounced, it goes through a second AlexNet to classify which kind of Phonological Process is found in
the word.
1 INTRODUCTION
The COVID-19 pandemic has brought many unin-
tended consequences for society, such as a nega-
tive effect on communication, especially in the pedi-
atric population, as preventative practices like wear-
ing masks, social distancing, and virtual classrooms
have a direct impact in the early-ages language de-
velopment since social interaction is essential for this
process.
Social Distancing and Virtual Classrooms have af-
fected children from having meaningful in-person in-
teractions with peers; this is a crucial component of
pragmatic development, which include conversational
skills as turn talking and understanding the meaning
behind the other speaker’s words. Masks also obscure
social cues provided by facial expressions (Charney
et al., 2021).
In 2021’s first quarter a survey was conducted by
the Education Endowment Foundation, which con-
sisted on collecting data from 58 primary schools
across England, showed that 76% of the pupils start-
ing school in September 2020 needed more communi-
cation support than in previous years; 96% of schools
claimed to be concerned about their pupils’ speech-
and-language development and 56% of parents were
concerned about their child starting school following
a
https://orcid.org/0000-0002-7510-618X
the lockdown in the spring and summer
1
.
Through the pandemic, parents have done an
amazing job to keep their children safe and healthy.
But this has reduced children’s exposure to new vo-
cabulary, as they are not able to hear and learn words
that family might use when the visit the extended fam-
ily. Mask wearing has also had an impact in language
development, as in school and pre-school, children
may struggle to differentiate between similar sounds
such as ”p” and ”t”, when their teacher is wearing a
mask. This can impact on a child’s speech develop-
ment or their phonological awareness; masks also ob-
scure facial expression, which contributes to how we
understand the meaning behind words
2
.
At the beginning of August 2020, the Peruvian
Ministry of Health presented a statement regarding
language and speech disorders during this pandemic.
In this statement, it was noted that children are one
of the most vulnerable population in the context we
live in, and there are reported cases of minors who
develop difficulty in articulating sounds, suffer alter-
ations in the fluency of speech and have problems in
the acquisition of expressive and / or comprehensive
language.
1
“Lockdowns hurt child speech and language skills -
report” - BBC
2
“How lockdown has affected children’s speech – and
what parents can do to help” - The Conversation
Baldeon, B., Ravelli, R. and Ugarte, W.
WawaSimi: Classification Techniques for Phonological Processes Identification in Children from 3 to 5 Years Old.
DOI: 10.5220/0010991300003182
In Proceedings of the 14th International Conference on Computer Supported Education (CSEDU 2022) - Volume 2, pages 147-154
ISBN: 978-989-758-562-3; ISSN: 2184-5026
Copyright
c
2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
147
The Ministry’s representative noted that confine-
ment causes children to have less social interaction,
which has a negative impact on increasing their vo-
cabulary and communication skills. She also noted
that by the end of 2020’s second quarter, Commu-
nity Mental Health Centers treated 6,846 cases of lan-
guage problems across the country
3
.
In order to address this problem, some researches
has been done, some of them on the classification of
mispronunciation of words, using different machine-
and-deep learning methods in order to do so. As an
example, for the Arabic alphabet, a comparison be-
tween classifiers was conducted; the results showed
that using a transfer-learning approach, taking ad-
vantage of an AlexNet outperformed other methods,
leading to a 99%+ accuracy on the task (Terbeh
and Zrigui, 2017). Due to this results, the transfer-
learning approach was selected in order to perform
the mispronunciation classification.
On the other hand, in order to conduct the identi-
fication of phonological processes in Brazil, consid-
ering different types and not only one, additional data
was provided by speech therapists (Franciscatto et al.,
2021). As we don’t have access to this data for Span-
ish, we decided to take the same approach as for the
mispronunciation classification.
Our contributions are as follows:
• We have built a synthetic dataset that mimics
both the pitch and speaking rate of 3-5 years old
Spanish-speaking children.
• We propose a neural network architecture to de-
tect mispronounced words based on their Mel-
spectrogram.
• We propose a neural network architecture to iden-
tify phonological processes (Structure of the syl-
lable and the word, Assimilation or Substitution)
in mispronounced words.
This paper is organized as follows. Section 2 dis-
cusses related work. Section 3 explains the context of
the problem and our main contributions in greater de-
tail. In section 4 we present the results obtained from
the experimentation and we conclude in section 5.
2 RELATED WORKS
In (Terbeh and Zrigui, 2017), the authors propose
the usage of 3 different neural networks (CNN, pre-
trained AlexNet and BLSTM) to classify which letter
of the Arabic alphabet has been pronounced and its
3
Ministry of Health warns of an increase in language
disorders in children due to the emergency (in spanish)
pronunciation quality. Additionally, they apply data
augmentation through pitch modification. After re-
moving noise and silences from each audio file, they
used Mel spectrograms, with three color channels,
as input data for the first two networks; and spec-
tral characteristics manually extracted for the last net-
work. Our proposal uses a pre-trained AlexNet to de-
tect bad pronunciation and PP’s; using MFCC heat
maps with a single color channel. On the other hand,
being in a scenario with no data available, we use the
modification of pitch and speaking rate to generate
synthetic data instead of just increasing it.
In (Nazir et al., 2019), The authors propose a
method based on characteristics extracted by a pre-
trained AlexNet and a method based on transfer learn-
ing to detect mispronunciation in people who learn
Arabic as a second language. In the first method, they
used the characteristics extracted by AlexNet from
Mel spectograms as input data from three different
models (KNN, SVM and a neural network). While in
the second method, they used the Mel spectrograms,
with three color channels, as input data for a pre-
trained AlexNet. Unlike them, we propose the use of
MFCC heat maps with a single color channel as input
data for a pre-trained AlexNet to detect mispronunci-
ation in children whose native language is Spanish.
In (Franciscatto et al., 2021), the authors presents
a method that consists on a Pronunciation Recog-
nition module, using a Decision Tree model, and a
Phonological Process Recognition module, that uses
a Random Forest model using a matrix given by an
expert with the probabilities of different Phonological
Processes in each phoneme of a set of words. Inspired
by this structure, our proposal uses a transfer-learnt
AlexNet to recognize mispronunciation and another
AlexNet for the identification of Phonological Pro-
cesses.
In (Shahnawazuddin et al., 2020), the authors
present a prosody-modification-based data augmen-
tation method for automatic speech recognition.
This is performed locating the Glottal Closure In-
stants (GCIs) to use them as anchors to modify the
Speaking-Rate and Pitch from the audios; after this
process, the audio is constructed again with the new
pitch and speaking-rate. While our method is very
similar, as we perform the modification of Speaking-
Rate and Pitch, we don’t calculate the GCIs, and mod-
ify the Speaking-Rate and Pitch directly.
CSEDU 2022 - 14th International Conference on Computer Supported Education
148
Table 1: Examples of errors (Pavez and Coloma, 2017).
(a) Syllable structure errors.
Structure of the syllable Error in spanish (Word in english)
Consonant-group reduction /p
´
ato/ for /pl
´
ato/ (plate)
Diphthong reduction /
´
a to/ for /
´
auto/ (car)
Coda suppression /pa tal
´
on/ for /pantal
´
on/ (pants)
Coalescence /k
´
en/ for /tren/ (train)
Omission of unstressed elements /p
´
osa/ for /mariposa/ (butterfly)
Addition of phonemes or syllables /n
´
ındio / for /indio/ (Indian)
Inversion of phonemes or syllables /u
´
ato/ for /
´
auto/ (car)
(b) Assimilation errors.
Assimilation Error in spanish (Word in english)
Identical bub
´
anda/ for /buf
´
anda/ (scarf)
Labial /pl
´
atamo/ for /pl
´
atano/ (banana)
Dental /mad
´
ıp
´
osa/ for /marip
´
osa/ (butterfly)
Velar /guf
´
anda/ for /buf
´
anda/ (scarf)
Nasal /anf
´
ombra/ for /a/f
´
ombra/ (carpet)
Syllabic /lilik
´
optero/ for /elik
´
optero/ (helicopter)
(c) Substitution errors.
Substitution Error in spanish (Word in english)
Posteriorization /ekif
´
ısio/ for /edif
´
ıcio/ (building)
Frontalization /bu
´
ante/ for /gu
´
ante/ (glove)
Stopping /p
´
oka/ for /f
´
oka/ (seal)
Fricativization /marif
´
osa/ for /marip
´
osa/ (butterfly)
Semiconsonantization of liquid phonemes /tjen/ for /tren/ (train)
Within-category liquid substitution /kape/usita/ for /kaperusita (Little Red Riding Hood)
3 IDENTIFYING AND
CLASSIFYING
PHONOLOGICAL PROCESSES
This section presents our method by introducing core
notions and developing them into a model proposal.
3.1 Preliminary Concepts
Now, we describe the main concepts for our proposal.
3.1.1 Phonological Processes
Also known as speech pattern errors, phonological
processes (PP’s) are mental operations that apply on
speech to substitute a class of sound or sound se-
quences that are difficult to the speech capacity of the
individual, for an alternative class identical but lack-
ing the difficulty property (Franciscatto et al., 2021).
Structure of the Syllable and Word. In the pro-
cesses related to the structure of the syllable and word,
WawaSimi: Classification Techniques for Phonological Processes Identification in Children from 3 to 5 Years Old
149
the emissions are simplified by bringing them closer
to the basic syllabic structure (consonant + vowel).
Some structure of the syllable and word phonologi-
cal processes are presented in Table 1 along with their
corresponding sub-class (Coloma et al., 2010).
Assimilation. Assimilation processes are strategies
by which a phoneme is replaced to make it the same or
similar to another present in the word (Coloma et al.,
2010). Some examples of this kind of phonological
processes are shown in Table 1b.
Substitution. In substitution processes, phonemes
belonging to one class are exchanged for members of
another phonemic class (Coloma et al., 2010). Some
examples of this kind of phonological processes are
shown in Table 1c.
3.1.2 Deep Learning
Deep learning uses a multi-layered cascade of non-
linear processing units for feature extraction and
transformation. The lower layers near the data en-
try learn simple features; in contrast, the upper layers
learn more complex characteristics derived from the
characteristics learned by the lower layer as shown in
Fig. 1. In recent years, deep learning has produced
cutting-edge results in many domains such as com-
puter vision, speech recognition, and NLP (Zhang
et al., 2018).
Figure 1: Characteristics learned by lower and upper lay-
ers (Zhang et al., 2018).
Convolutional Neural Network. Is a deep learning
model used to analyze and learn from data or infor-
mation that has a mesh or grid pattern. These models
generally have three different types of layers: convo-
lutional, grouping, and fully connected. The first two
perform the feature extraction process autonomously
through convolutions and the third locates these fea-
tures in the final output. A convolution is a type of
linear operation used to extract features; basically, a
small matrix of numbers, called a kernel, is applied to
the input data, tensor (a matrix of numbers) as shown
in Fig 2. The tensor and the kernel are used to cal-
culate a scalar product for each element of the ten-
sor; the result is used as the output value in the cor-
responding position of the characteristic map (Khan
et al., 2020).
Figure 2: Convolution example (Khan et al., 2020).
Transfer Learning: This method aims to provide a
framework for using previously acquired knowledge
to solve new but similar problems much more quickly
and effectively. This method has been inspired by the
fact that human beings can use previously acquired
knowledge to solve new but similar problems; for ex-
ample, learning to recognize apples can help us to rec-
ognize pears, or learning to play the electronic organ
Figure 3: Workflow for applying transfer learning with
CNN (Lu et al., 2015).
CSEDU 2022 - 14th International Conference on Computer Supported Education
150
Table 2: Target words to identify phonological processes.
Correctly pronounced Incorrectly pronounced
/alf
´
ombra/ (carpet) /a f
´
ombra/
/
´
auto/ (car) /
´
a to/, /u
´
ato/
/buf
´
anda/ (scarf) /bub
´
anda/, /guf
´
anda/
/kaperusita/ (Little Red Riding Hood) /kaperutusita/, /kape usita/
/edif
´
ıcio/ (building) /ekif
´
ısio/
/f
´
oka/ (seal) /p
´
oka/
/gu
´
ante/ (glove) /bu
´
ante/
/elik
´
optero/ (helicopter) /lilik
´
optero/
/indio/ (Indian) /n
´
ındio/
/mariposa/ (butterfly) /p
´
osa/, /mad
´
ıp
´
osa/, /marif
´
osa/
/pantal
´
on/ (pant) /pa tal
´
on/
/pl
´
atano/ (banana) /pl
´
atamo/
/pl
´
ato/ (plate) /p
´
ato/
/tel
´
efono/ (phone) /ten
´
efolo/
/tren/ (train) /t en/, /k
´
en/, /tjen/
Table 3: Number of generated audio files.
Correctly pronounced Incorrectly pronounced Total
Male (10) 10 × 15 × 6 = 900 (10 × 22 − 1) × 6 = 1, 314 2,214
Female (5) 5 × 15 × 6 = 450 5 × 22 × 6 = 660 1,110
Total (15) 1,350 1,974 3,324
can make learning the piano easier (Lu et al., 2015).
The way we apply this method in a convolutional net-
work context is shown in Fig 3.
3.1.3 Voice Conversion
Voice Conversion (VC) is an area of speech process-
ing that deals with the conversion of the speaker’s per-
ceived identity. VC aims to modify the non-linguistic
or paralinguistic information of speech while preserv-
ing the linguistic information; that is, the voice signal
emitted by a first speaker, the source speaker, is mod-
ified so that it sounds as if it were spoken by a sec-
ond speaker, the target speaker (Popa et al., 2012) as
shown in Fig. 4.
3.2 Method
In this section we will explain our main contributions:
creation of a synthetic data using voice conversion,
Figure 4: Voice conversion example (Wu and Li, 2014).
applying transfer learning in a CNN to detect mispro-
nounced words and the same process to identify PP’s.
3.2.1 Synthetic Dataset
Due to the fact that, to the best of our knowledge,
there was not a corpus of Spanish children’s speech
available so far, we had to use Voice Conversion to
generate synthetic audios that imitate the voice of 3-
to-5 years old children.
This process consists of modifying the pitch and
speaking rate of audios collected from people be-
tween 20 and 25 years old. In order to perform
this modification, we considered an average pitch of
309 Hz for 3-years-old children (Schuckman, 2008),
WawaSimi: Classification Techniques for Phonological Processes Identification in Children from 3 to 5 Years Old
151
(a) Mispronunciation detection process workflow. (b) Phonological process identification workflow.
Figure 5: Processes Workflows.
255Hz for 4 years and 253 Hz for 5 years (Cappellari
and Cielo, 2008). Likewise, we use a speaking rate of
5 or 7 phonemes per second for all three ages (Shah-
nawazuddin et al., 2020), resulting in 6 combinations.
A group of 15 persons were asked to record their
voices while pronouncing a set of words commonly
used to identify Phonological Processes (Popa et al.,
2012). This set of words consist on 15 words, and 22
mispronunciation of this 15 words as seen in Table 2;
giving a total of 37 audios.
After collecting the audios, they were processed
detecting and removing silences from the beginning
and end of them in order to get the audios with only
the words; these processed files were modified (pitch
and speaking rate) using the Python librosa library.
From this process we obtained a synthetic corpus of
3324 audio files, because one person only recorded 36
words, as shown in the Table 3.
3.2.2 Mispronunciation Detection
To detect whether a word was pronounced correctly
or incorrectly, we used the Mel frequency cepstral co-
efficients (MFCC) of the generated audio files and
a previously trained convolutional neural network
(AlexNet). We use AlexNet available in the PyTorch
library, which was trained on the ImageNet dataset
and can classify between 1000 different classes.
The detail of our sequence of steps is shown in
Fig. 5. Firstly, we extract the MFCC using the librosa
library and resize its heat map to 227 × 227 × 1. Sec-
ondly, we import the Alexnet and modify the first con-
volutional layer, such that it has the same input size as
the heat maps, and the last two layers completely con-
nected, so that it has an output according to our num-
ber of classes (correctly and incorrectly pronounced).
Thirdly, we divided the corpus into training (70%),
test (30%) and validation (222 audios); for which we
employ two approaches ”Known Speaker” and ”Un-
known Speaker”.
• Known Speaker (KS): the partition is done with-
out taking into account the source of the audio. In
such a way that the audios emitted by the same
person can be in the training and test partitions.
• Unknown Speaker (US): the partition is carried
out taking into account the source of the audio. In
such a way that the audios emitted by the same
person will only be in one of the partitions.
Finally, we carry out the training and testing of the
network. We train using the same parameters as (Ter-
beh and Zrigui, 2017), changing the batch size to 4
and using 20 and 30 epochs.
3.2.3 Phonological Processes Identification
As can be seen in Fig. 5, the process to identify
phonological processes is very similar to that de-
scribed in Fig. 5. The only differences are the input
and output data; that is, the AlexNet in charge of this
process is trained only with the mispronounced words
of the corpus and its completely connected layers are
modified based on the tree classes of PP’s (Structure
of the syllable and word, Assimilation and Substitu-
tion).
4 EXPERIMENTS
In this section we will explain the experiments we car-
ried out to see the performance of our solution.
4.1 Experimental Protocol
All of the experiments were performed using a
Google Collaborate notebook. The free tier gives us a
CSEDU 2022 - 14th International Conference on Computer Supported Education
152
Table 4: Results obtained by the models.
(a) Mispronunciation detection models.
Accuracy Precision Recall F-measure
Baseline (Franciscatto et al., 2021) 92.5% - - -
Our Proposal (KS-20 epochs) 85.3% 83.1% 94.8% 88.6%
Our Proposal (KS-30 epochs) 87.8% 89.9% 89.5% 89.7%
Our Proposal (US-20 epochs) 82.8% 87.0% 83.4% 85.2%
Our Proposal (US-30 epochs) 93.4% 93.5% 95.4% 94.5%
(b) Phonological processes identification models.
Accuracy Precision Recall F-measure
Baseline (Franciscatto et al., 2021) 94.6% - - -
Our Proposal (KS-20 epochs) 97.6% 98.3% 93.6% 95.8%
Our Proposal (KS-30 epochs) 99.1% 98.7% 98.1% 98.4%
Our Proposal (US-20 epochs) 89.6% 87.6% 78.5% 82.8%
Our Proposal (US-30 epochs) 94.7% 95.6% 86.7% 90.6%
(c) Comparison of trained models with and without voice conversion
Accuracy Precision Recall F-measure
Model trained with adult voices 77.0% 74.5% 93.1% 82.9%
Model trained with generated voices 79.8% 82.2% 84.0% 83.1%
virtual machine with a single-core Intel Xeon, 13GB
of Ram and a 32GB HDD.
We used the programming language Python v.3.6.
As well as the librosa v.0.8.1, scikit-image v.0.16.2
and PyTorch v.1.9.0+cu111 libraries.
Our dataset and code are freely available at https:
//github.com/Senzouz/WawaSimi
4.2 Results
As we used a different methodology than in (Fran-
ciscatto et al., 2021), we compared our Deep Learn-
ing proposal with the Machine Learning proposal pre-
sented in (Franciscatto et al., 2021). We trained our
model using the combination of 20 and 30 epochs
with both data-split approaches, Known Speaker and
Unknown Speaker. The results of our experiments in
terms of the mispronunciation detection model can be
seen in Table 4.
In the case of the Phonological Processes Iden-
tification proposal, we performed the same experi-
ments as in the Mispronunciation Detection method
and compared it with the highest accuracy model re-
ported in (Franciscatto et al., 2021). This results are
shown in Table 4b.
Inspired by (Shahnawazuddin et al., 2020), we de-
cided to replicate their experiments but with our meth-
ods to validate the effectiveness of the voice conver-
sion. This means, we train a model with the 554 orig-
inal audios (adult voices) and another model with the
same amount of generated audios, which were ran-
domly selected. Both models were tested with 222
generated audios selected for validation, the results
obtained can be viewed in Table 4c.
4.3 Discussion
As can be seen in Table 4, our best model obtained
93.4% accuracy when detecting mispronunciation.
This is probably because the model was trained with
the Unknown Speaker, which is better suited for new
instances than Known Speaker as talkers are only part
of the training or testing phase, instead of being ran-
domly distributed between both phases.
On the other hand, as shown in Table 4b, our
best model obtained 99.1% accuracy when identify-
WawaSimi: Classification Techniques for Phonological Processes Identification in Children from 3 to 5 Years Old
153
ing PP’s. This may happen because in this process the
acoustic characteristics of a voice are more influential.
Therefore, the Known Speaker approach obtains bet-
ter results because the model trains with all the voices
available in the dataset learning their acoustic charac-
teristics.
As can be seen in the Table 4c, the voice con-
version method improves the performance of models
when detecting mispronunciation. When we trained
the model with only 554 synthetic audios the accuracy
metric increased 2.8%. However, when using the full
set of synthetic audios as in the Table 4, the perfor-
mance increased by 26.4%.
5 CONCLUSIONS AND
PERSPECTIVES
Thanks to the usage of a transfer-learnt AlexNet, we
were able to develop a method to detect mispronunci-
ation in Spanish as well as identify phonological pro-
cesses with 93.4% and 99.1% respectively.
While language-learning apps like Duolingo,
Babbel also employ classification methods to deter-
mine whether a word is partially or fully mispro-
nounced (de la Cal Rioja, 2016). We carry out a sec-
ond classification to determine the type of phonologi-
cal process present in the misspronunciation.
Furthermore, machine and deep learning tech-
niques are being used for various voice recognition
tasks (e.g., query by humming (Alfaro-Paredes et al.,
2021)), that could lead us to improvements in our ap-
proach. Additionally, text formalization (de Rivero
et al., 2021) could be applied for educational purposes
in compliment with our proposal.
For future works, we would like to have the sup-
port of a speech and language therapist who will help
us with the construction of a data set that includes
as many possible mispronunciation scenarios. In this
way, we reduce the probability that the models do not
know how to classify a new audio instance.
REFERENCES
Alfaro-Paredes, E., Alfaro-Carrasco, L., and Ugarte, W.
(2021). Query by humming for song identification us-
ing voice isolation. In IEA/AIE.
Cappellari, V. M. and Cielo, C. A. (2008). Vocal acoustic
characteristics in pre-school aged children. Brazilian
Journal of Otorhinolaryngology, 74(2).
Charney, S. A., Camarata, S. M., and Chern, A. (2021). Po-
tential impact of the covid-19 pandemic on communi-
cation and language skills in children. Otolaryngol-
ogy–Head and Neck Surgery, 165(1).
Coloma, C. J., Pavez, M. M., Maggiolo, M., and Pe
˜
naloza,
C. (2010). Desarrollo fonol
´
ogico en ni
˜
nos de 3 y 4
a
˜
nos seg
´
un la fonolog
´
ıa natural: Incidencia de la edad
y del g
´
enero. Revista signos, 43(72).
de la Cal Rioja, J. (2016). Hound word. software
para la mejora de la pronunciaci
´
on en ingl
´
es.
Universidad de Valladolid - Undergraduate thesis.
https://uvadoc.uva.es/handle/10324/17963.
de Rivero, M., Tirado, C., and Ugarte, W. (2021). For-
malstyler: GPT based model for formal style trans-
fer based on formality and meaning preservation. In
KDIR.
Franciscatto, M. H., Fabro, M. D. D., Lima, J. C. D., Trois,
C., Moro, A., Maran, V., and Soares, M. K. (2021).
Towards a speech therapy support system based on
phonological processes early detection. Comput.
Speech Lang., 65.
Khan, A., Sohail, A., Zahoora, U., and Qureshi, A. S.
(2020). A survey of the recent architectures of deep
convolutional neural networks. Artif. Intell. Rev.,
53(8).
Lu, J., Behbood, V., Hao, P., Zuo, H., Xue, S., and Zhang,
G. (2015). Transfer learning using computational in-
telligence: A survey. Knowl. Based Syst., 80.
Nazir, F., Majeed, M. N., Ghazanfar, M. A., and Maq-
sood, M. (2019). Mispronunciation detection using
deep convolutional neural network features and trans-
fer learning-based model for arabic phonemes. IEEE
Access, 7.
Pavez, M. M. and Coloma, C. J. (2017). Phonological
problems in spanish-speaking children. Advances in
Speech-language Pathology.
Popa, V., Sil
´
en, H., Nurminen, J., and Gabbouj, M. (2012).
Local linear transformation for voice conversion. In
ICASSP. IEEE.
Schuckman, M. (2008). Voice characteristics of preschool
age children. PhD thesis, Miami University.
Shahnawazuddin, S., Adiga, N., Kathania, H. K., and Sai,
B. T. (2020). Creating speaker independent ASR sys-
tem through prosody modification based data augmen-
tation. Pattern Recognit. Lett., 131.
Terbeh, N. and Zrigui, M. (2017). Identification of pro-
nunciation defects in spoken arabic language. In PA-
CLING, volume 781 of Communications in Computer
and Information Science.
Wu, Z. and Li, H. (2014). Voice conversion versus speaker
verification: an overview. APSIPA Transactions on
Signal and Information Processing, 3.
Zhang, L., Wang, S., and Liu, B. (2018). Deep learning for
sentiment analysis: A survey. Wiley Interdiscip. Rev.
Data Min. Knowl. Discov., 8(4).
CSEDU 2022 - 14th International Conference on Computer Supported Education
154
