Detecting Dyslexia from Audio Records: An AI Approach

Jim Radford, Gilles Richard, Hugo Richard and Mathieu Serrurier

Dystech Ltd, Traralgon, Australia

Keywords:

Dyslexia, Screening, Machine Learning.

Abstract:

Dyslexia impacts the individual’s ability to read, interferes with academic achievements and may also have

long term consequences beyond the learning years. Early detection is critical. It is usually done via a lengthy

battery of tests: human experts score these tests to decide whether the child requires speciﬁc education strate-

gies. This human assessment can also lead to inconsistencies. That is why there is a strong need for earlier,

simpler (and cheaper) screening of dyslexia. In this paper, we investigate the potential of modern Artiﬁ-

cial Intelligence in automating this screening. With this aim in mind and building upon previous works, we

have gathered a dataset of audio recordings, from both non-dyslexic and dyslexic children. After proper pre-

processing, we have applied diverse machine learning algorithms in order to check if some hidden patterns

are discoverable, making a difference between dyslexic and non-dyslexic readers. Then, we built up our own

neural network which outperforms the other tested approaches. Our results suggests the possibility to clas-

sify audio records as characteristic of dyslexia, leading to an accurate and inexpensive dyslexia screening via

non-invasive methods, potentially reaching a large population for early intervention.

1 INTRODUCTION

Learning disorders such as dyslexia, dyspraxia, dys-

graphia, etc., are deeply connected to an individual’s

outcomes not only during their academic years, but

also when it comes to employment, mental health

and more. Despite the fact there is no exact ﬁgure

about the number of dyslexic people on earth, it is

widely accepted by the community that dyslexia af-

fects about 5%-10% of school-age children and, if we

include adults, then it can go up to 15% (see the re-

port coming from Duke University (Duke, 2016) for

instance). The Diagnostic and Statistical Manual of

Mental Disorders also known as DSM–5 (American

Psychiatric Association, 2013), is often considered as

one of the reference documents on this matter. To go

to the point, Dyslexia is:

• deﬁned as a basic deﬁcit in learning to read (i.e.

decode print) (Vellutino et al., 2004),

• characterized by a signiﬁcant impairment in the

development of reading skills,

• observable by reading performances well below

the normal range for given age groups and IQ lev-

els (L. et al., 2016),

• not explained by sensory deﬁcits such as visual,

hearing impairment, insufﬁcient scholarship or

overall mental development only (L. et al., 2016).

At least, the scientiﬁc community agrees on this def-

inition even if some technical details remain debat-

able (Waesche et al., 2011). Clarifying the causes

of dyslexia has been a serious goal of research over

the past twenty years. Despite much progress be-

ing made across diverse ﬁelds, the causes of dyslexia

still remain opaque and there is no scientiﬁc consen-

sus. Several theories coexist, some of them have al-

ready been discredited by empirical observations, oth-

ers still remain as serious candidates waiting for con-

ﬁrmation. Testing these hypotheses is a difﬁcult task:

dyslexic people do not form a homogeneous popu-

lation and exhibit diverse patterns of errors. That

is also why dyslexia is often divided into subtypes

(phonological, visual, etc.), possibly originating from

deﬁcits at various stages of the comprehension sys-

tem. Nevertheless, following (Tamboer et al., 2014),

it is possible to identify dyslexia with a high reliabil-

ity, although the exact nature of dyslexia is still un-

known.

Dyslexia remains throughout a persons lifetime

but can be mitigated with appropriate training ses-

sions. Obviously, it is not a matter of Yes or No and

the symptoms range from mild to severe.

Without known causes, detecting dyslexia remains

a challenging task. The process is complex and per-

Radford, J., Richard, G., Richard, H. and Serrurier, M.

Detecting Dyslexia from Audio Records: An AI Approach.

DOI: 10.5220/0010196000580066

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 58-66

ISBN: 978-989-758-490-9

 2021 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reserved

formed by an accredited professional. This profes-

sional looks for many different indicators, intended

to detect whether reading, writing (and also calcu-

lus skills) are being acquired at a proper rate. Gen-

erally a questionnaire to acquire information regard-

ing medical history, social environment, school per-

formances, etc. is ﬁlled out, then the ﬁnal diagnos-

tic is done via a battery of tests (mainly paper-based).

The questionnaire is mainly a self-report which, in the

case of a child, can be supervised by the tutor. The

assessment may be lengthy, costly and emotionally

painful. Moreover the limited number of accredited

professionals may make this process time consuming.

Without mentioning the cost of the assessment which

can also be prohibitive.

Very often, dyslexic children can get government

supports in diverse ways (speciﬁc teaching lessons,

extra tuition, extra time for exams, speciﬁc staff help-

ing during the classroom, etc.). To get such a support,

the criteria is to provide a certiﬁcate coming from an

accredited specialist. This situation obviously pre-

vents a lot of people, often among the most vulnera-

ble members of the population, to carry on an assess-

ment. This makes the search of easy, fast, reliable and

widely available assessments (or pre-assessments) of

primary interest.

From another perspective, there has been an in-

creasing interest to carry over the success stories

of Artiﬁcial Intelligence (AI) and Machine Learning

(ML) in diverse domains, from natural language un-

derstanding to cancer screening through autonomous

vehicles. That is why some researchers from the AI

community have started to investigate ML techniques

for dyslexia screening

In this paper, our initial assumption is that a prop-

erly trained ML algorithm might be able to distin-

guish between non dyslexic children and dyslexic

children, starting from a set of simple features. In the

case of dyslexia, we consider features extracted from

audio records of word reading (known words and non-

sense words

). Our experiences on a restricted set

of data coming from speech pathologist partners in

Australia demonstrate the soundness of this approach.

This paper is organized as follows. In Section 2,

we discuss existing related works. In Section 3, we

describe the main principles underlying our method.

Section 4 is dedicated to our experiments: we de-

scribe the dataset, the protocol and the results we get.

Word screening is more appropriate at this stage. But, in

the remainder of this paper, we indifferently use the words

diagnostic, assessment and screening.

A nonsense word follows phonetic rules, has no meaning

and does not belong to the English dictionary but looks like

a proper English word.

Before providing future works in our conclusion Sec-

tion 6, we discuss the limitations of our current ap-

proach in Section 5.

2 RELATED WORKS

We are not alone in our belief that Machine Learning

or AI in the large part are there to help provide an

effective dyslexia screening. Among the recent (post

2010) AI-based approaches to diagnose dyslexia, we

can roughly distinguish 2 categories:

- Category 1: Approaches using the results of

human-expert scoring to provide a diagnostic. In

these cases, the diagnostic process does not change a

lot: the user has still to undertake a battery of tests and

these tests are human-marked. Since it may be dif-

ﬁcult to maintain children’s attention throughout the

tests, an option is to include all the tests in a serious-

game, to better grab the attention of the participant.

But only the ﬁnal diagnostic is done via an AI algo-

rithm.

- Category 2: Approaches taking, as an initial

assumption, one of the candidate theories explain-

ing dyslexia. In these cases, the AI algorithm is fed

with data related to the underlying theory. In the

case of neurological explanations for dyslexia, the au-

thors use brain scans or EEG. In case of oculo-motor

deﬁcits origin, the authors implement eye-tracking

technologies.

2.1 Multi-tests Approaches (Cat. 1)

The approaches considered in this section take as in-

put the results of tests undertaken by the user. These

tests can be:

• pre-deﬁned tests generally administered by an ac-

credited professional. In this case, the tests can

also be marked by the practitioner. Optionally

a computer-assisted marking process is imple-

mented.

• speciﬁcally designed computer-based tests captur-

ing information relevant to dyslexia. In this case,

a professional is needed to supervise a child par-

ticipant.

Then the output of these tests are fed into an AI al-

gorithm which ultimately provides a diagnostic as a

likelihood of dyslexia (i.e. a number between 0 and

1). We can cite the works of (Palacios et al., 2010;

Costa et al., 2013; Al-Barhamtoshy and Motaweh,

2017; Shamsuddin et al., 2017) which are still at a

research stage. Only the work of (Rello et al., 2018),

Detecting Dyslexia from Audio Records: An AI Approach

implemented via a serious game, led to a commer-

cial system Dytective

. The authors suggest an accu-

racy of 85% on their dataset, which is relatively good.

Another interesting option has been recently imple-

mented in (Spoon et al., 2019), (which is not category

1, stricto sensu). Because ‘reading is intimately con-

nected to writing, and children who struggle to learn

to read often struggle to write’, Spoon et al. provide

a proof-of-concept by developing a system that used

computer vision to classify handwriting samples as

indicative of dyslexia or not. They get 77.6% accu-

racy in determining whether a patch of handwriting

was written by a student with dyslexia or not. Still

in early development, these works provide relatively

good results using very accessible information with

no need for a questionnaire or battery of tests.

2.2 Serious Games Approaches (Cat. 1)

When one wants to target a population of very young

children (let’s say under 7), it becomes crucial to de-

sign a data gathering process which is sufﬁciently at-

tractive to motivate them. Serious games are therefore

natural candidates for this task. An early attempt has

been done by (Lyytinen et al., 2007), but it was before

the effective emergence of AI. The works of (Van den

Audenaeren et al., 2013) within the DYSL-X project

and the works of (Gaggi et al., 2012) are targeted to

this aim: they have designed serious games, dedicated

to young children, available on computers or tablets

and allowing the measurement of some parameters

characteristic of dyslexia. These works are more fo-

cused on the design and implementation of a system

than on estimating its accuracy. The work of (Geurts

et al., 2015) DIESEL-X was also developed to detect

a high risk for having dyslexia in preschoolers. Sev-

eral theories on the underlying cause of dyslexia are

converging on the idea that one fundamental problem

derives from abnormal neurological timing, or ’tem-

poral processing’ (Johnson, 1980). As a consequence,

the perception of musical elements could differ from

children with and without dyslexia. Starting from

this assumption, (Rauschenberger et al., 2017) pro-

poses DysMusic, a prototype which aims to predict

the risk of having dyslexia before acquiring reading

skills. The prototype was designed to observe par-

ticipants listening to music (via a web app), using

the think aloud protocol, varying different acoustic

parameters such as frequency, duration which relate

with perceptual parameters such as pitch and loud-

ness. Among other tasks, the participants (10 in the

cohort) have to evaluate how difﬁcult it was to dis-

Dytective is a cross-platform app available on

https://www.changedyslexia.org.

tinguish between the various sounds. (Gaggi et al.,

2017) have developed a set of 6 serious games, using

a 2D graphic design, and experimented on 24 partici-

pants. Both (Rauschenberger et al., 2017) and (Gaggi

et al., 2017) have experimented on very small cohorts

but their works still demonstrate that serious gaming

is an interesting pathway for dyslexia screening.

2.3 Neuro-based Approaches (Cat. 2)

Neuro-based works start from the widely admitted as-

sumption that dyslexia is linked to a speciﬁc brain

conﬁguration, either in terms of anatomical shape

or in terms of functional organisation. In (Tamboer

et al., 2016) three dimensional whole-brain scans are

acquired from each participant, each acquisition se-

quence lasting approximately 6 minutes. Using a

standard classiﬁer, they properly classify 80% of the

scans between dyslexics and non dyslexics. This ac-

curacy declines to 59% with a larger range of par-

ticipants. In fact, their algorithm provides a large

percentage of false alarms, i.e. many people with-

out dyslexia are labelled with dyslexia (which is a

cautious behavior but can lead to serious stress for a

child).

The authors of (Frid and Manevitz, 2018) start

from the assumption (Asynchrony Theory) that

dyslexia could come from a gap in the speed of pro-

cessing between the different brain entities activated

in the word decoding process. This gap may pre-

vent the synchronization of information necessary for

an accurate reading process. Starting from this, they

monitor a population of 32 children, with a more or

less 50/50 percentage of dyslexic/non dyslexic read-

ers. Getting the children to read 96 real words and

96 non-sense words, they record brain activity via

electroencephalogram (EEG) and implement a binary

classiﬁcation algorithm (namely Support Vector Ma-

chines SVM) to distinguish between dyslexic and non

dyslexic readers. They obtain an accuracy in the range

of 78.5%. Obviously, this kind of technique cannot

be used in a realistic way to screen a large part of the

population. And, due to the heavily controlled en-

vironment needed to capture the data, it is unlikely

these neuro-based approaches will lead to a publicly

available tool anytime soon.

2.4 Eye Tracking-based Approaches

(Cat. 2)

Eye-tracking can provide serious insight into percep-

tual/cognitive processes (PH. et al., 2013). Despite

the fact that the work of (Hyona et al., 1995) tends

to dismiss the oculo-motor dysfunction hypothesis of

HEALTHINF 2021 - 14th International Conference on Health Informatics

dyslexia, it is a fact that today most studies agree that

there is a link between visual-attention and oculo-

motor control during reading: see for instance (Bel-

locchi et al., 2012; Huettig and Brouwer, 2015) for

recent publications on this topic. From an AI perspec-

tive, it is then natural to monitor the eye movement of

a user during reading activity.

For the ﬁrst time in (Rello and Ballesteros, 2015),

an eye tracking technology associated to an SVM

classiﬁer was used to predict dyslexia starting from

a dataset of 97 subjects, 48 of them with diagnosed

dyslexia. The eye tracking technology allows the ex-

traction of information such as Number of visits (To-

tal number of visits to the area of interest in the text),

Mean of visit (Duration of each individual visit within

the area of interest in the text), etc. The resulting ac-

curacy is in the range of 80% which is relatively good.

(M. et al., 2016) start from the same idea but moni-

tor different parameters such as for a given eye sac-

cade, the duration of the event, the distance spanning

the event, the average eye position during the event,

etc. Their standard classiﬁer shows a very high ac-

curacy (around 96%). The whole process also takes

into account the results from a battery of other com-

mon tests, such as rapid automatic naming, reading

of nonsense words, etc.: this obviously increases the

duration of a screening session

More recently, we can cite the works of

(Asvestopoulou et al., 2019), still using eye tracking

associated to an SVM classiﬁer and getting an excel-

lent 97% accuracy rate over a set of 69 native Greek

speaking children, 32 being dyslexic. Their system

DysLexML, still a work in progress, could ultimately

be the basis of another screening tool. Nevertheless,

eye tracking-based methods still need an external de-

vice to be connected in one way or another to a com-

puter. This could be considered as a serious draw-

back.

Our work departs from all the previous ones in

terms of the data we use, in terms of simplicity and in

terms of duration of a session. In the following sec-

tion, we describe how we proceed to tackle the issue

of dyslexia screening.

3 PREDICTION PRINCIPLE

Because one of the main symptoms of dyslexia is dif-

ﬁculty in reading, we have decided to only gather

reading audio recordings, from both dyslexic and non

Starting from this approach, Lexplore

(https://www.lexplore.com/) has been founded in 2016 in

Sweden which then expanded to the USA in 2017.

dyslexic readers, then to apply machine learning al-

gorithms. Instead of analysing images, brain signals

or eyes movements, we directly analyse audio signals.

We agree that poor reading performance is not an ulti-

mate marker of dyslexia, but our results demonstrate

that a dedicated machine learning algorithm associ-

ated with proper audio signal processing can extract

patterns that are not accessible to a human expert. Let

us start with what a user is supposed to provide.

3.1 Word Selection and Generation

Our process is to have every individual to read 32

words (no sentences, only words). It is well-known

that dyslexic children struggle when it comes to read-

ing words they have never seen or heard. They have

also difﬁculties with some letters, or combinations of

letters (p and q for instance) and certain syllables.

Our initial corpus comes from a set of 82 children’s

books extracted from the Gutenberg Project (Hart,

1971). We clean the texts and remove proper nouns.

We obtain a list of around 100 000 words. Then we

produce two lists : one with words from 4 to 6 letters,

one with words from 7 to 9 letters. In each list, we

consider only words with a high frequency of occur-

rence to guarantee the words are known by children.

After ﬁltering, each of the two lists contains around

2000 words.

In a second step, we create two lists of nonsense

words. We also need to guarantee that the nonsense

is pronounceable. In order to achieve that, we build a

Long-Short Term Memory neural networks (LSTM)

(Hochreiter S, 1997) that learn to build such nonsense

words. We are then able to generate an inﬁnite list of

nonsense words

. As for the real words, we build two

lists of nonsense words with different size (1000 non-

sense words with 4 to 6 letters, 2500 nonsense words

with 7 to 9 letters) and we only keep nonsense words

that ﬁt within the following constraints :

• Every subset of 4 consecutive letters exists in

an English word (to guarantee the word is pro-

nounceable)

• It contains difﬁcult letters or a difﬁcult combina-

tion of letters for dyslexic people.

The ﬁnal list of 32 words to be read by a child

is obtained by choosing 16 words in the list of real

words and 16 words in the list of nonsense words. We

change the length of the words according to the age

of the user performing the assessments:

• List 1: From 6 to 8 years old (included) the list is

ordered this way:

We also use the expression ’generated words’ because they

are the output of an AI process.

Detecting Dyslexia from Audio Records: An AI Approach

– 2 easy real words

–

of words from 4 to 6 letters (50% real, 50%

nonsense):

–

of words from 7 to 9 letters (50% real, 50%

nonsense):

• List 2: From 9 to 13 years old (included) the list

is ordered this way:

– 2 easy real words

–

of words from 4 to 6 letters (50% real, 50%

nonsense):

–

of words from 7 to 9 letters (50% real, 50%

nonsense):

• List 3: 14 years old and over the list is ordered

this way:

– 2 easy real words

– 100% of words from 7 to 9 letters (50% real,

50% nonsense):

These constrained lists of words are randomly

generated and are age-related: short words with sim-

ple syllables for children from 7 to 8, more difﬁcult

for children from 9 to 13, then difﬁcult for children

over 14. It is then very unlikely that 2 sessions lead

to the same list of 32 words

. Note that 50% of the

words are displayed with the times new roman font

and 50% are displayed with the Open Dyslexic font.

We tune the size of the font to ensure that the vowels

appears with exactly the same size on the screen.

3.2 Input Parameters for Dyslexia

Classiﬁer

For every audio record, we consider 2 parameters:

• The Reading Reaction Time (RRT) is the inter-

val between the initial display of the word and the

start of the reading.

• The Reading Time (RT) is the time it takes for the

user to read the corresponding word.

In both RRT and RT, the time unit is millisecond (ms)

and their evaluation is done via a computer (no human

in the loop). Consequently, from a session of 32 audio

records, we extract 6 numbers which will be used in

our ML experiments:

• average RRT for 32 words, average RRT for 16

real words, average RRT for 16 nonsense words

• average RT for 32 words, average RT for 16 real

words, average RT for 16 nonsense words

In fact, recently, we have decided to put at the end of the

lists, 2 easy real words: it is better to ﬁnish a session on a

positive!

Note: In an ideal word, RRT

real

+ RRT

gene

= 2 ×RRT ,

but it often happens that some words are not read by

the user. Then this simple linear relation is not valid

anymore.

For instance, a simple analysis of the reaction time

(RRT) reading time (RT) on our dataset is given in

Figure 1. This demonstrates that dyslexic readers

need more time than non dyslexic ones, whatever the

type of words to be read.

Figure 1: Reaction time (left table) - reading time (right

table) for non-dyslexic/dyslexic readers w.r.t. the type of

word.

We assessed 93 users, among them 43 are dyslexic

and 50 are non dyslexic readers. All in all, we

achieved 93 x 32= 2976 audio records. Due to this

small cohort, it is not realistic to consider a sim-

ple threshold approach such as RRT > T hresRRT

and RT > T hresRT would be enough to characterize

dyslexic readers. Our ﬁnal neural network computes

a more sophisticated formula.

4 EXPERIMENTS

We provide in the following subsections the general

context of screening dyslexia in a global population,

the precise metrics we use to be in a position to rig-

orously compare different algorithms and the results

we get from each of these algorithms. Everything

has been implemented in Python, with Tensorﬂow and

SciKit Learn libraries.

4.1 Context

Machine learning is a data-driven technology. Apart

from designing an algorithm, we have to gather

proper data, with their own labels. Not only the qual-

ity, but also the quantity is important since training

machine learning algorithms may require a lot of data

in order to be accurate. In (Wagner, 2018), the au-

thor points out that the difﬁculty to diagnose dyslexia

is mainly coming from the unbalanced population.

Only a relatively low percentage of the population has

dyslexia or dysgraphia, and so to train a mathematical

model with such an unbalanced population is always

a challenge. When it comes to measuring the perfor-

HEALTHINF 2021 - 14th International Conference on Health Informatics

mance of the algorithm, standard accuracy can then

be a misleading metric. Assuming we have 10% of

the population having dyslexia or dysgraphia, a base-

line algorithm declaring everybody as non dyslexic

will ensure a 90% accuracy.

4.2 Metrics

As a consequence, other metrics are needed such as

precision, recall and F1-score. Let us recall below

some standard deﬁnitions. For a binary classiﬁer,

we have at our disposal a set of positive examples

(dyslexic children) and a set of negative examples

(non dyslexic children). We note t p the number of

positive examples predicted as positive, tn the num-

ber of negative examples predicted as negative, f p

the number of negative examples predicted as posi-

tive (false positive), and f n the number of positive

examples predicted as negative (false negative). The

metrics are deﬁned as follows:

accuracy =

t p + tn

t p + tn + f p + f n

(1)

The accuracy measures the probability that the class

predicted by the model is the right one. In the latter

we will use percentage for this accuracy.

precision =

t p

t p + f p

(2)

The precision is the probability of being positive if

the example is predicted as positive. In some sense,

this measures the correctness of the predictor when

it predicts an example as positive. The bigger this

number, the better the predictor is.

recall =

t p

t p + f n

(3)

The recall is the probability of a positive example to

be predicted as positive. In some sense, this measures

the ability of the predictor to predict all positive ex-

amples as positive. Still, the bigger this number, the

better the predictor is.

F1-score = 2 ∗

precision ∗ recall

precision + recall

(4)

The F1-score is a balance between precision and re-

call. Thus, accuracy focus on the performances of the

model in general when precision, recall and F1-score

focus on performances of the model on positive ex-

amples only.

4.3 Results

First of all, our baseline is what we call the Dummy

classiﬁer that chooses classes randomly with the a pri-

ori probability (their frequency) of classes computed

on the training set: this provide an accuracy of 0.46

which is quite poor. Then, we compare with the per-

formances of state of the art classiﬁers: Logistic Re-

gression (LR), KNN, SVM with polynomial kernel

(SVC), SVM with linear kernel (LSVC), Naive Bayes

NB, Random Forest (RFC) and Decision Tree (DTR).

As we just want an estimation of their related per-

formance, we use the default parameters for each of

these classiﬁers, using scikit-learn Python library. We

are aware that spending time to tune these parameters

could lead to better results.

In a second step, we build up a neural network NN

with 4 dense layers (activation function Relu) then a

ﬁnal dense with sigmoid activation function as it is

usual for binary classiﬁcation.

In order to have an average estimation of the met-

rics previously described:

1. We perform for each classiﬁer a 10-fold cross-

validation scheme and we get the above 4 metrics.

2. We repeat the experiences 10000 times and our

table provides the average values of the 4 metrics

on these 10000 runs.

Table 1: Performance of different machine learning algo-

rithms for detecting dyslexia from audio signals.

Algorithm Accuracy Precision Recall F1

Logistic reg. 69.00 0.72 0.63 0.64

KNN 69.01 0.73 0.65 0.66

ıve Bayes 69.02 0.73 0.58 0.61

Poly. SVM 68.10 0.76 0.33 0.43

Linear SVM 69.00 0.71 0.62 0.63

Random Forest 68.04 0.75 0.65 0.67

Decision Tree 68.01 0.64 0.66 0.62

Neural Network 81.72 0.86 0.72 0.78

The average values of metrics are shown in Table

1 (in all the tables, the subtitle ’dys’ means dyslexic

reader, ’no dys’ means non dyslexic reader). The best

results are obtained by the neural network and the

other classiﬁers (when they are not tuned) are more

or less equivalent in terms of performance.

The network, providing an accuracy of more than

80% could probably be tuned to get better perfor-

mances. But due to the relatively small number of

data that we have, this could lead to over ﬁtting with-

out giving a clear picture of the accuracy in a real

environment. This could be partially overcome by

considering more data (provided they are not biased).

Getting more data is part of our future task. How-

ever, these experiences clearly demonstrate the power

of an ML approach to distinguish dyslexic from non

dyslexic readers.

Detecting Dyslexia from Audio Records: An AI Approach

Figure 2: Error rate for non-dyslexic/dyslexic readers w.r.t.

type of word (left table) and type of font (right table).

Figure 3: Reaction time (left table) and reading time (right

table) for non-dyslexic/dyslexic readers w.r.t. the font.

4.3.1 Additional Experiments

At this stage of our work, we have introduced other

features that we tag manually. For instance, for each

spoken word in the data set, we manually indicate

if the reading was right or not. Figure 2 highlights

that the error rate is signiﬁcantly higher for dyslexic

people (which was expected). Also, the nonsense

words are more difﬁcult to read for dyslexic than for

non-dyslexic people, consequently generating more

errors: the gap in error rate between dyslexic and non

dyslexic is bigger for nonsense words than for real

words. The error rate seems to be a good indicator of

dyslexia: we use it in our machine learning approach.

By considering the error rate, we achieve an ac-

curacy of 90% on our dataset (in the same conditions

than the previous experiments). This clearly estab-

lishes the interest of the error rate for improving the

quality of the screening. Because this error rate is not

automatically detected, it is not taken into account in

Table 1. We currently work on machine learning ap-

proaches to compute this error rate automatically.

We also consider the type of font we use to dis-

play the word on the screen. Figure 2 shows that the

use of the Open Dyslexic font decreases the error rate

for both dyslexic and non dyslexic. Nevertheless, we

do not observe from Figure 3 a similar positive effects

for reaction time and reading time. The improvement

of accuracy with Open Dyslexic font has already been

reported in (de Leeuw, 2010) but this is still a debat-

able issue (Wery and Diliberto, 2017). The two last

studies also report no effect on the reaction and read-

ing time, which is in accordance with what we got. Fi-

nally, we have tested that including the reaction time

with respect to the font as a new input feature for the

neural network does not improve its accuracy.

5 LIMITATIONS AND OTHER

OPTIONS

Despite our promising results, we are well aware that

we have to be cautious.

• First of all, despite our dataset being generally

bigger than the ones of our competitors, it is too

small to provide a deﬁnitive conclusion. The sizes

of dataset used in modern machine learning tech-

nologies are far bigger than a hundred or so exam-

ples.

• If more data are needed, it is quite clear that we

need data coming from a controlled environment

to be sure of their accuracy. Having noisy data

will degrade our prediction.

• Spending time in tuning the standard algorithms

(SVM, random forests, etc.) could lead to better

results that the ones we got. One thing is for sure:

building a basic neural network such as the one

we use is effortless and immediately brings better

results than non-tuned standard algorithms.

• We are currently focusing on getting an automatic

pronunciation error detection which would deﬁ-

nitely improve our accuracy. But why not using

other parameters such as total reading time (in-

stead of an average value) which is widely used

in the Gray Oral Reading Test aka GORT 5 for

instance ((Wiederholt and Bryant, 2012)).

6 FUTURE WORKS AND

CONCLUSION

Analysing data (verbal/written test results, brain im-

ages, eyes movement, etc.) via Machine Learning

technologies to detect dyslexia is not a new idea. In

this paper, we show that it is also possible to predict

dyslexia by analysing audio signals using ML. Our

method satisﬁes the requirements needed to build a

mass market screening tool:

• We focus on the human observable symptoms of

dyslexia,

• We do not use any other data than the audio

records,

• We do not use any external device to gather data,

• A screening session is between 10 to 15 minutes

long.

It has been recently proven that properly trained ML-

based predictors can be more accurate than human ex-

perts on speciﬁc tasks. Based on these facts and our

HEALTHINF 2021 - 14th International Conference on Health Informatics

encouraging results, we think there is a huge poten-

tial for ML-based technologies to help people with

dyslexia. As usual with ML, accuracy can still be

improved by gathering more data. ML-based tech-

nologies could deﬁnitely avoid the need of manual

analysis and global performances may be improved.

In the future, a better understanding of the correla-

tion between the different disorders could also help

in providing more informed predictions. For instance

adding to audio records, a picture of a handwritten

text could help to make the prediction still more ac-

curate. As far as we know, this is the ﬁrst time audio

signals analysis is used to detect dyslexia, leading the

path to a non invasive, fast and cost effective screen-

ing tool.

REFERENCES

Al-Barhamtoshy, H. M. and Motaweh, D. M. (2017). Di-

agnosis of dyslexia using computation analysis. In

2017 International Conference on Informatics, Health

Technology (ICIHT), pages 1–7.

American Psychiatric Association (2013). Diagnostic and

statistical manual of mental disorders: DSM-5. Autor,

Washington, DC, 5th ed. edition.

Asvestopoulou, T., Manousaki, V., Psistakis, A., Smyr-

nakis, I., Andreadakis, V., Aslanides, I. M., and

Papadopouli, M. (2019). Dyslexml: Screening

tool for dyslexia using machine learning. CoRR,

abs/1903.06274.

Bellocchi, S., Mathilde, M., Bastien-Toniazzo, M., and

Ducrot, S. (2012). I can read it in your eyes: What eye

movements tell us about visuo-attentional processes in

developmental dyslexia. Research in developmental

disabilities, 34:452–460.

Costa, M., Zavaleta, J., Cruz, S., Manhaes, L., Cerceau,

R., Carvalho, L., and Mousinho, R. (2013). A com-

putational approach for screening dyslexia. In Pro-

ceedings of the 26th IEEE International Symposium

on Computer-Based Medical Systems, pages 565–566.

de Leeuw, R. (2010). Special font for dyslexia?

Duke, U. (2016). Dyslexia international: Bet-

ter training, better teaching. https://www.

dyslexia-international.org/wp-content/uploads/

2016/04/DI-Duke-Report-ﬁnal-4-29-14.pdf.

Frid, A. and Manevitz, L. M. (2018). Features and machine

learning for correlating and classifying between brain

areas and dyslexia. CoRR, abs/1812.10622.

Gaggi, O., Galiazzo, G., Palazzi, C., Facoetti, A., and

Franceschini, S. (2012). A serious game for predicting

the risk of developmental dyslexia in pre-readers chil-

dren. In 21st International Conference on Computer

Communications and Networks (ICCCN), pages 1–5.

Gaggi, O., Palazzi, C. E., Ciman, M., Galiazzo, G., Frances-

chini, S., Rufﬁno, M., Gori, S., and Facoetti, A.

(2017). Serious games for early identiﬁcation of de-

velopmental dyslexia. Comput. Entertain., 15(2):4:1–

4:24.

Geurts, L., Vanden Abeele, V., Celis, V., Husson, J., Au-

denaeren, L., Loyez, L., Goeleven, A., Wouters, J.,

and Ghesqui

ere, P. (2015). DIESEL-X: A game-based

tool for early risk detection of dyslexia in preschool-

ers, pages 93–114. Springer International Publishing.

Hart, M. (1971). Project gutenberg. www.gutenberg.org.

Hochreiter S, S. J. (1997). Long short-term memory. Neural

Computation, 9(8):1735–1780.

Huettig, F. and Brouwer, S. (2015). Delayed anticipatory

spoken language processing in adults with dyslexia-

evidence from eye-tracking: Word reading and pre-

dictive language processing. Dyslexia, 21.

Hyona, J., Olson, R., Defries, J., Fulker, D., Penning-

ton, B., and Smith, S. (1995). Eye ﬁxation patterns

among dyslexic and normal readers: Effects of word

length and word frequency. Journal of Experimen-

tal Psychology: Learning, Memory, and Cognition,

21:1430–1440.

Johnson, D. J. (1980). Persistent auditory disorders in

young dyslexic adults. Bulletin of the Orton Society,

30:268–276.

L., V., F., M., G., T., and M., H. (2016). Direct view-

ing of dyslexics’ compensatory strategies in speech

in noise using auditory classiﬁcation images. PLoS

ONE, 11(4).

Lyytinen, H., Ronimus, M., Alanko, A., Poikkeus, A.-

M., and Taanila, M. (2007). Early identiﬁcation of

dyslexia and the use of computer game-based prac-

tice to support reading acquisition. Nordic Psychol-

ogy, 59(2):109–126.

M., N. B., G., O. S., J, Y., A., P. T. R., and C, J. (2016).

Screening for dyslexia using eye tracking during read-

ing. PLoS ONE, 11(12).

Palacios, A. M., S

anchez, L., and Couso, I. (2010). Diag-

nosis of dyslexia with low quality data with genetic

fuzzy systems. International Journal of Approximate

Reasoning, 51(8):993 – 1009.

PH., T., IG., C., G., P., JN., R., DP., M., and L., I. (2013).

High-throughput classiﬁcation of clinical populations

from natural viewing eye movements. J Neurol.,

1(260):275–284.

Rauschenberger, M., Rello, L., Baeza-Yates, R., Gomez, E.,

and Bigham, J. P. (2017). Towards the prediction of

dyslexia by a web-based game with musical elements.

In Proceedings of the 14th Web for All Conference on

The Future of Accessible Work, W4A ’17. Association

for Computing Machinery.

Rello, L. and Ballesteros, M. (2015). Detecting readers

with dyslexia using machine learning with eye track-

ing measures. Proceedings of the 12th Web for All

Conference W4A ’15, pages 1–8.

Rello, L., Romero, E., Rauschenberger, M., Ali, A.,

Williams, K., Bigham, J. P., and White, N. C. (2018).

Screening dyslexia for english using HCI measures

and machine learning. In Kostkova, P., Grasso, F.,

Castillo, C., Mejova, Y., Bosman, A., and Edel-

stein, M., editors, Proceedings of the 2018 Interna-

tional Conference on Digital Health, DH 2018, Lyon,

France, April 23-26, 2018, pages 80–84. ACM.

Detecting Dyslexia from Audio Records: An AI Approach

Shamsuddin, S. N. W., Mat, N. S. F. N., Makhtar, M.,

and Isa, W. M. W. (2017). Classiﬁcation tech-

niques for early detection of dyslexia using computer-

based screening test. World Applied Sciences Journal,

35(10).

Spoon, K., Crandall, D., and Siek, K. (2019). Towards de-

tecting dyslexia in children’s handwriting using neural

networks. In ICML Workshop on AI for Social Good.

Tamboer, P., Vorst, H. C. M., Ghebreab, S., and Scholte,

H. S. (2016). Machine learning and dyslexia: Clas-

siﬁcation of individual structural neuro-imaging scans

of students with and without dyslexia. NeuroImage.

Clinical, 11:508–514.

Tamboer, P., Vorst, H. C. M., and Oort, F. J. (2014). Identi-

fying dyslexia in adults: an iterative method using the

predictive value of item scores and self-report ques-

tions. Annals of Dyslexia, 64(1):34–56.

Van den Audenaeren, L., Celis, V., Vanden Abeele, V.,

Geurts, L., Husson, J., Ghesqui

ere, P., Wouters, J.,

Loyez, L., and Goeleven, A. (2013). Dysl-x: Design

of a tablet game for early risk detection of dyslexia in

preschoolers. In Games for Health, pages 257–266.

Springer Fachmedien Wiesbaden.

Vellutino, F., Fletcher, J., Snowling, M., and Scanlon, D.

(2004). Speciﬁc reading disability (dyslexia): what

have we learned in the past four decades? Journal

of child psychology and psychiatry, and allied disci-

plines, 45 1:2–40.

Waesche, J., Schatschneider, C., Maner, J., Ahmed, Y., and

Wagner, R. (2011). Examining agreement and longi-

tudinal stability among traditional and rti-based deﬁ-

nitions of reading disability using the affected-status

agreement statistic. Journal of learning disabilities,

44:296–307.

Wagner, R. K. (2018). Why is it so difﬁcult to diag-

nose dyslexia and how can we do it better? https:

//dyslexiaida.org/.

Wery, J. J. and Diliberto, J. A. (2017). The effect of a spe-

cialized dyslexia font, opendyslexic, on reading rate

and accuracy. Annals of dyslexia, 67(2):114–127.

Wiederholt, J. L. and Bryant, B. R. (2012). (GORT-5) Gray

Oral Reading Test, Fifth Edition. Pro-Ed.

HEALTHINF 2021 - 14th International Conference on Health Informatics