SEVERE APNOEA DETECTION USING SPEAKER

RECOGNITION TECHNIQUES

Rubén Fernández, Jose Luis Blanco, Luis A. Hernández, Eduardo López

GAPS: Signal, Systems & RadioComm. Department Universidad Politécnica de Madrid, Spain

José Alcázar

Respiratory Department. Unidad de Trastornos Respiratorios del Sueño Hospital Clínico Universitario Málaga, Spain

Doroteo T. Toledano

ATVS Biometric Recognition Group. Universidad Autonoma de Madrid, Spain

Keywords: Apnoea, Automatic Speaker Recognition techniques, GMM, Nasalization.

Abstract: The aim of this paper is to study new possibilities of using Automatic Speaker Recognition techniques

(ASR) for detection of patients with severe obstructive sleep apnoea (OSA). Early detection of severe

apnoea cases can be very useful to give priority to their early treatment optimizing the expensive and time-

consuming tests of current diagnosis methods based on full overnight sleep in a hospital. This work is part

of an on-going collaborative project between medical and signal processing communities to promote new

research efforts on automatic OSA diagnosis through speech processing technologies applied on a carefully

designed speech database of healthy subjects and apnoea patients. So far, in this contribution we present and

discuss several approaches of applying generative Gaussian Mixture Models (GMMs), generally used in

ASR systems, to model specific acoustic properties of continuous speech signals in different linguistic

contexts reflecting discriminative physiological characteristics found in OSA patients. Finally, experimental

results on the discriminative power of speaker recognition techniques adapted to severe apnoea detection are

presented. These results obtain a correct classification rate of 81.25%, representing a promising result

underlining the interest of this research framework and opening further perspectives for improvement using

more specific speech recognition technologies.

1 INTRODUCTION

Obstructive sleep apnoea (OSA) is a highly

prevalent disease (Puertas et al., 2005), affecting an

estimated 2-4% of population between the ages of

30 and 60 years. It is characterized by recurring

episodes of upper airway obstruction during sleep,

and it is usually associated with loud snoring and

increased daytime sleepiness. OSA is a serious

threat to an individual’s health if not treated (Puertas

et al., 2005). OSA is a risk factor for cardiovascular

disease (Coccagna et al., 2006), it is usually related

to traffic accidents caused by somnolent drivers

(Puertas et al., 2005; Lloberes et al., 2000), and it

can lead to a poor quality of life and impaired work

performance. A full overnight sleep study is usually

required to diagnose OSA; referred to as

polysomnography it involves the recording of

neuroelectrophisiological and cardiorespiratory

variables. Polysomnography is an expensive and

time-consuming test so patients usually have to

suffer a waiting list of several years before the test is

done (Puertas et al., 2005). At present, the most

effective and widespread treatment for OSA is nasal

CPAP (Continuous Positive Airway Pressure) which

prevents apnoea episodes providing a pneumatic

splint to the airway.

In this article we investigate the acoustical

characteristics of patients’ speech with OSA with the

purpose of relying on an automatic system for severe

apnoea voice assessment. The acoustic properties of

speakers suffering from obstructive sleep apnoea are

124

Fernández R., Blanco J., A. Hernández L., López E., Alcazar J. and T. Toledano D. (2009).

SEVERE APNOEA DETECTION USING SPEAKER RECOGNITION TECHNIQUES.

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 124-130

DOI: 10.5220/0001546601240130

 SciTePress

not clear, and not much research has been made on

it. However, there are some studies pointing at

possible abnormalities in phonation, articulation and

resonance (Fox & Monoson, 1989). In a previous

research, we designed and collected an apnoea

speech database (Fernández et al., 2008) based on

both, the review of previous research in this field,

and a preliminary manual contrastive study

performed on a small group of healthy subjects and

apnoea patients. This contrastive study helped us to

identify phonetically related acoustic characteristics

of patients’ speech with OSA.

In this work, we explore the possibilities for the

adaptation of Automatic Speaker Recognition

techniques (ASR) to severe apnoea cases diagnosis.

To our knowledge this represents a pioneer research

on automatic severe OSA diagnosis using signal

processing algorithms on continuous speech. The

proposed method intends to be complementary to

existing OSA diagnosis methods (i.e.

Polysomnography) and clinician perceptual

judgments, to help in an early detection of these

cases. Early severe OSA detection can enable more

efficient medical protocols, giving higher priority to

more risky cases (there is a high risk of car accident

due to somnolence in severe apnoea patients) and

thus optimizing both social benefits and medical

resources. The proposed system is based on

Gaussian Mixture Models (GMMs), which

represents the state-of-the-art for speaker recognition

(Reynolds et al., 2000).

More specifically, our work is mainly focused on

the nasality factor, since it has been traditionally

identified as an important feature in the acoustic

characteristics of apnoea speakers. The term

nasalization can refer to two different phenomena in

the context of speech; hyponasality and

hypernasality. The first type occurs if no

nasalization is produced when the speech sound

should actually be nasal. This is the case when for

example having a cold. Hypernasality is nasalization

during the production of non-nasal (voiced oral)

sounds. The interested reader can find an excellent

reference work in (Pruthi, 2007). Up to now no

signal processing and pattern recognition techniques

have been applied to the analysis hyponasal or

hypernasal continuous speech in OSA patients,

therefore this will represent one of the aims of our

research.

The rest of this paper is organized as follows:

Section 2 addresses the physiological and acoustic

characteristics in OSA patients. The speech database

used in our experimental work is described in

Section 3. Section 4 explores the possibilities of

using GMM approaches to study and model

nasalization in OSA patient’s speech. Application of

speaker recognition technologies for automatic

severe apnoea detection is presented in Section 5.

Finally, some conclusions and future research are

given in Section 6.

2 PHYSIOLOGICAL &

ACOUSTIC

CHARACTERISTICS IN OSA

SPEAKERS

Nowadays is still not clear neither the

articulatory/physiologically settings nor the acoustic

characteristics of speech in apnoea speakers. Most of

the more valuable information in this field can be

found in Fox and Monoson work (Fox & Monoson,,

1989) where a perceptual study with skilled judges

was presented comparing voices from apnoea

patients and a control group (referred to healthy

subjects). This study showed that although

differences between both groups of speakers were

found, acoustic cues for these differences are

somewhat contradictory and unclear. What seemed

to be clear was that apnoea group had abnormal

resonances that might be due to altered structure or

function of the upper airway, and theoretically, this

anomaly should result not only in a respiratory but

also in a speech dysfunction. Consequently, the

presence of speech disorder in a OSA population

should be expected, and it could include

abnormalities in articulation, phonation and

resonance:

• Articulatory Anomalies: Fox and Monoson

stated that neuromotor dysfunction could be

found in a sleep apnoea population as a “lack of

regulated innervations to the breathing

musculature or upper airway muscle hypotonus”

This dysfunction is normally related to speech

disorders, especially dysarthria. There are

several types of dysarthria, resulting in various

different acoustic features. All types of

dysarthria affect the articulation of consonants

and vowels causing the slurring of speech.

Another common feature in apnoea patients is

hypernasality and problems with respiration.

• Phonation Anomalies: The phonation

anomalies may be due to the fact that the heavy

snoring of sleep apnoea patients can cause the

inflammation in the upper respiratory system and

affect the vocal cords.

SEVERE APNOEA DETECTION USING SPEAKER RECOGNITION TECHNIQUES

125

• Resonance Anomalies: The analysis of

resonance characteristics for the sleep apnoea

group in (Fox & Monoson, 1989) did not yield

any clear conclusion. These authors only

conclude that resonance abnormalities in apnoea

patients could be perceived as hyponasality or

hypernasality. It is only recently that resonance

disorder affecting speech sound quality has been

associated with vocal tract damping features

distinct from airflow in balance between the oral

and nasal cavities. The term applied to this

speech disorder is “cul-de-sac” resonance and is

used for a type of hyponasality.

2.1 Initial Contrastive Acoustic Study

A manual contrastive study of an initial version of

our apnoea speech database was made. A similar

revision can be found in (Fiz et al, 1993) where an

acoustic spectral analysis was applied to sustained

vowels to detect possible OSA cases. Nevertheless,

this study doesn’t investigate all the possible

acoustic characteristics in apnoea patients’ voice

because the use of sustained vowels excludes the

study of other acoustic clues present in continuous

speech and measures based on the comparison

between different linguistic contexts (Parsa &

Jamieson, 2001)..

For our preliminary study, a group of 16

speakers, eight in the apnoea group and eight in the

control group, were recorded. All the speakers

uttered a same group of 25 phonetically balanced

sentences; they were extracted from the phonetic

corpus of Albayzin database (Moreno et al., 1993), a

resource for speech recognition technology in

Spanish. The contrastive study was performed by

visual comparison of frequency representations

(mainly spectrographic, pitch, energy and formant

analysis) of apnoea and control group speakers.

Following some conclusions of previous research

on acoustic characteristics of apnoea speakers we

checked nasality feature. The existence and the size

of one extra low frequency formant was studied

(Glass & Zue, 1985) as indicator of nasalization, but

no perceptual differences between the groups could

be found. As discussed in (Fox & Monoson, 1989),

this fact could be explained by the perceptual

difficulty in classifying the apnoea speakers’ voice

as hyponasal or hypernasal disorder. However,

studies on specific linguistic contexts were also

made and differences in the variation in nasalization

degree between different linguistic contexts (nasal

and non-nasal context) of both groups were found,

there being a tendency for the apnoea speakers to

have smaller differences between those contexts.

This is indeed very interesting. A hypothesis has

been that apnoea speakers might have an overall

higher nasality level due to velopharyngeal

dysfunction, and that the differences between oral

would be small, because the oral vowels would also

be nasalized. One explanation could be that the

control of the velopharyngeal mechanism is lower

for the apnoea speakers, which could lead to

difficulties in making changes of nasality, although

the absolute nasalization might be either high or low

Another interesting conclusion of our study was,

related to resonance anomalies, when we compare

the distance between the second and third formant

for the /i/ vowel, clear differences between the

apnoea and control groups were found. Apnoea

speakers showed a higher distance and this was

especially clear in diphthongs where /i/ was stressed,

for example in the Spanish word “Suiza” ('suj θa).

According to (Hidalgo & Quillis, 2002) the position

of the third formant might be related to the degree of

velopharyngeal opening. Lowering of velum gives

rise to higher frequencies of the third formant.

3 APNOEA DATABASE

The database collection was performed in the

Respiratory Department at Hospital Clínico

Universitario of Málaga, Spain. The database was

composed of about 80 male subjects; half of them

with severe sleep apnoea (AHI > 30), and the other

half healthy subjects or with mild OSA (AHI < 10).

Subjects in both groups have similar physical

characteristics such as age and body mass index.

Furthermore, speech material for the apnoea group

was recorded and collected in two different sessions:

one just before being diagnosed and the other after

several months under CPAP treatment. The analysis

of speech characteristics in these two sessions will

allow studying the evolution of apnoea patient’s

voice characteristics after the treatment.

3.1 Speech and Image Collection

Speech signal was recorded with a sampling

frequency of 48 kHz in an acoustically isolated

booth. Recording equipment was a standard laptop

computer equipped with a SP500 Plantronics

headset microphone that includes A/D conversion

and digital data exchange through USB-port.

Additionally, for each subject in the database,

two facial images (frontal and lateral views) were

collected under controlled illumination conditions

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

126

and over a white flat background. A conventional

digital camera was used to obtain images in a 24-bit

RGB format, without compression and with a

2272x1704 resolution. We decide to collect visual

information because OSA is associated to a variable

combination of different anatomic factors and a

visual exam of the patient is suggested as first step

in the examination of a subject with clinical

suspicion of OSA (Puertas et al., 2005), valuing the

morphotype (i.e. obesity, short neck, etc.) and facial

features (i.e. malocclusions referred to bad bites,

etc.). To our knowledge, no research has been made

on finding evidences that could relate facial features

extracted through image processing of patients’ face

pictures to severe apnoea cases diagnosis. Thus we

decided to include both frontal and lateral pictures of

each patient’s face making it possible to research on

this challenging topic.

3.2 Speech Corpus

The speech corpus consists in four sentences

repeated three times and a sustained vowel /a/ also

repeated three times. The four sentences were

designed trying to cover all the relevant linguistic

contexts where physiological OSA features could

have higher impact on specific acoustic

characteristics of particular phonemes. Particular

emphasis has been put on:

• Sentence design that allows intra-speaker

variation measurements; that is to measure

differential voice features for each speaker, as,

for example, the nasalization degree between

vowels inside and outside nasal contexts.

• Continuous voiced sounds to measure irregular

phonation patterns related to muscular fatigue of

apnoea patients.

• Voiced sound in the context of guttural

phonemes to analyse the specific impact of

possible articulatory dysfunctions in the

pharyngeal region.

• A broad range of vowel sounds to allow the

accurate modelling of the articulation space and

other measures related to our preliminary

contrastive study.

All sentences where designed to exhibit a similar

melodic structure, and speakers where requested to

read them following this rhythmic structure. We

hope this controlled rhythmic recording procedure

could help in avoiding non-relevant inter-speaker

linguistic variability. Each sentence is presented in

the following paragraphs where the melodic groups

have been underlined.

1. Francia, Suiza y Hungría

ya hicieron

causa común. ('fraN θja 'suj θa i uŋ 'gri a

ya j 'θje roŋ 'kaw sa ko 'mun).

2. Julián no vio la manga roja

que ellos

buscan, en ningún almacén.(xu 'ljan no

'βjo la 'maŋ ga 'řo xa ke 'e λoz 'βus kan en

niŋ 'gun al ma 'ken).

3. Juan no puso la taza rota

que tanto le

gusta en el aljibe. (xwan no 'pu so la 'ta θa

'řo ta ke 'taN to le 'γus ta en el al 'xi βe).

4. Miguel y Manu llamarán entre ocho y

nueve y media. (mi 'γel i 'ma nu λa ma 'ran

'eN tre 'o t∫o i 'nwe βe i 'me ðja).

First phrase includes stressed vowels /i/ in

diphthongs where (according to our preliminary

contrastive study, see 2.1) differences between the

second and third formant could be distinctive in

apnoea speakers. Second and third phrases, both

negatives, have similar grammatical and intonation

structure. They could be useful to make contrastive

studies of vowels in different linguistic contexts.

Some examples of these contrastive pairs are: a

nasal context, “manga roja” ('maŋ ga 'řo xa) versus

a neutral context, “taza rota” ('ta θa 'řo ta). As we

mention in previous sections, these contrastive

analyses could be indeed very interesting to confirm

that apnoea speakers might have an overall higher

nasality level and smaller intra-speaker differences

between non-nasal and nasal vowels due to

velopharyngeal dysfunction. Fourth phrase includes

a relatively long sentence mainly composed of

voiced sounds read as one single melodic group in

order to discover possible phonation abnormalities

during the sustained generation of voiced sounds.

The three repetitions of the sustained vowel /a/,

including onset and offset, could be used to study

phonation measures (i.e. jitter and shimmer) that

detect the variations in amplitude and frequency of

the speakers, but in a more controlled way.

4 GMM FOR OSA VOICE

MODELLING

In this section we present an initial experimentation

directed to evaluate the capabilities of using

Gaussian Mixture Models techniques (GMM) to

model perceptual acoustic characteristics described

in previous research (Fox & Monoson, 1989) and

corroborated in our initial contrastive study.

In particular, as our speech database was

designed to allow a detailed contrastive analysis of

SEVERE APNOEA DETECTION USING SPEAKER RECOGNITION TECHNIQUES

127

vowels in oral and nasal phonetic contexts, we

focussed on reporting perceptual differences related

to resonance anomalies and perceived as

hyponasality or hypernasality. To that purpose,

GMM-based approach techniques were applied to

confirm these differences in nasalization degree

between different linguistic contexts, as was already

described in Section 2 and was revealed in the initial

contrastive study.

4.1 GMM Training and Testing

Protocol

Gaussian Mixture Models (GMMs) and adaptation

algorithms are effective and efficient pattern

recognition techniques suitable for sparse speech

data modelling in Automatic Speaker Recognition

systems (Reynolds et al., 2000) that we will apply

for apnoea voice modelling. In our experimental

framework, the ASR open source tool BECARS

(Blouet et al., 2004) was used. Details on

parameterization, model training and classification

phases for the BECARS baseline system are the

following:

• Parameterization consists in extracting

information from speech signal (in our case, 39

MEL frequency cepstral coefficients –MFCC-,

delta MFCCs and delta-delta MFCCs).

• GMMs for apnoea and control groups were

trained as follows. First, a universal background

model (UBM) was trained on a subcorpus of

Albayzin Database (Moreno et al., 1993), thus

entirely separated from the apnoea speech

database. Next, speech from apnoea and control

group speakers was used to train the apnoea and

control group GMM models. Both models were

trained using MAP adaptation from the UBM.

Only means were adapted (as classically done in

speaker verification). This technique increases

robustness of the models, especially when little

speech material is available. Obviously, speech

data from speakers for the model training was

not included in the test set.

For testing purposes, in order to increase the

number of tests and thus to improve the statistical

relevance of our results, the standard leave-x-out

testing protocol was used. This protocol consists in

discarding x speakers from the experimental set,

training some models on the remaining data and

testing with that x speakers. This scheme is

replicated until reaching a sufficient number of tests.

4.2 OSA Resonance Anomalies Study

using GMMs

With the aim of testing the capabilities of using

GMM technique to perform contrastive analysis on

perceived resonance anomalies, two GMMs for each

apnoea and healthy speaker were modelled using

speech from nasalized and non-nasalized vowels.

Both, nasal and non-nasal GMM models were

trained using a similar approach described in section

4.1; MAP adaptation from a generic vowel UBM

model trained using Albayzin. These two nasal/non-

nasal GMM models were used to quantify the

acoustic differences between nasal and non-nasal

contexts for both apnoea and control groups. The

smaller difference between nasal/non-nasal GMM

models, the more similar are both nasalized and non-

nasalized vowels; thus revealing the presence of

resonance anomalies. As distance measure between

nasal/non-nasal GMM models, a fast approximation

of the Kullback-Leibler (KL) divergence for

Gaussian Mixture Models was used (Do, 2003). This

distance is used in ASR to define cohorts or groups

of speakers producing similar sounds.

As can be seen in Figure 1, significant

differences in the nasal/non-nasal GMM distances

were found for control and apnoea groups. This

result clearly shows that acoustic differences

between oral and nasal vowels are smaller in apnoea

speakers and confirms the trend to an overall higher

nasality level revealed in previous research. As a

result, GMM approach appears as an appropriate and

suitable speech technique for further research on

modelling and detecting severe apnoea speakers.

Figure 1: Kullback distances differences between

nasal and non-nasal models for apnoea and control

speakers.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

128

5 GMM-BASED SPEAKER

RECOGNITION TECHNIQUES

FOR SEVERE APNOEA

DETECTION

As we have seen in the previous section, GMM

approach is able to model some of the already

described resonance abnormalities perceived in

apnoea speakers. Now, this experimental set-up was

intended to test the potential of applying standard

ASR techniques to the automatic diagnosis of

apnoea disease. A speaker verification system is a

supervised classification system able to discriminate

speech signals into two classes (genuine and

impostor). In our case, the classes will not

correspond to a given speaker, but a sleep apnoea

class and a control class (referred to healthy

subjects). This method is suitable for keeping track

the evolution of voice dysfunction among the

patients, is easy-to-use, quick, non-invasive for the

patients and affordable for the clinicians. The

proposed method is not expected to replace the

existing OSA diagnosis methods, but to help in an

early detection of severe apnoea cases.

Following a similar approach already used for

other pathological voice assessment studies

(Fredouille et al., 2005), GMM models representing

the apnoea and control classes were built as follows:

• The pathological and control GMM models

were trained from the generic UBM model

thanks to the MAP adaptation technique using

standard leave-x-out technique, in a similar

method as described above (Section 4).

• During the apnoea/non-apnoea detection phase,

an input speech signal corresponding to the

utterance of a speaker to be diagnosed is

presented to the system. The parameterised

speech is then processed with each apnoea and

control GMM model generating two likelihood

scores. From these two scores an

apnoea/control decision is made.

In these initial experiments the task of control and

apnoea voice classification has been investigated.

Table 1 provides the correct classification rates

obtained. An overall correct classification of 81.25

% is reached on this task.

Table 1: Correct Classification Rate.

Correct

Classi.

Rate in %

Control

Group

Apnoea

Group

Overall

77,50 %

(31/40)

85 %

(34/40)

81,25 %

(65/80)

Table 2 shows the statistical measures of the

performance for the same task (Fisher's exact test

p<0.0001 revealed statistically significant). Results

are presented in terms of Sensitivity: proportion of

positives which are correctly identified as such;

Specificity: proportion of negatives which are

correctly identified; Positive Predictive Value:

proportion of patients with positive test results who

are correctly diagnosed; Negative Predictive Value:

proportion of patients with negative test results who

are correctly diagnosed.

Table 2: Sensitivity, Specificity, Positive Predictive Value

and Negative Predictive Value.

Sensitivity Specificity

Positive

Predictive

Value

Negative

Predictive

Value

77,50 %

(31/40)

85 %

(34/40)

83,78 %

(31/37)

79,06 %

(34/43)

Considering the correct classification rates

achieved, there are some aspects to comment. The

results are encouraging and it seems that apnoea

information may be caught by a GMM based-

approach, even with few training speech. These

promising results have been obtained without

particular attention to the choice of acoustic

parameters used, so best results can be expected with

a representation of the audio data optimized for

pathology discrimination. Evidently, these

experiments have to be validated with a higher

population, but the results have already given us a

preliminary notion of the discriminative power of

this approach for the automatic diagnosis of severe

apnoea cases.

6 CONCLUSIONS AND FUTURE

RESEARCH

In this paper the acoustic properties of speakers with

severe obstructive sleep apnoea (OSA) and the

adaptation of standard ASR techniques for detection

and modelling of these patients have been presented.

This study represents a pioneer research in the field

SEVERE APNOEA DETECTION USING SPEAKER RECOGNITION TECHNIQUES

129

of automatic diagnosis of OSA disease.. Preliminary

experimental results on the speech database

collected using state-of-the-art GMM speaker

recognition techniques shows that it is worth

continuing the research on this area. Related to

nasality factor as an important feature in the acoustic

characteristics of apnoea speakers, GMM approach

have confirmed that there are significant differences

between apnoea and control group on the relative

nasalization degree between different linguistic

contexts. Therefore, future research will be focused

on exploiting this information in order to use it for

the automatic apnoea diagnosis. Furthermore, best

results can be expected with a representation of the

audio data optimized for pathology discrimination

On the other hand, and bearing in mind the

speech database design criteria, we propose the use

of other acoustic measures usually applied over

pathological voices (jitter, HNR, etc.). These

techniques could also be applied over different

linguistic and phonetic contexts, and could be fussed

to GMM approach to improve our initial

discrimination results.

ACKNOWLEDGEMENTS

The activities described in this paper were funded by

the Spanish Ministry of Science and Technology as

part of the TEC2006-13170-C02-01 project. The

authors would like to thank the volunteers at

Hospital Clínico Universitario of Málaga, Spain, and

to Guillermo Portillo who made the speech and

image data collection possible.

REFERENCES

Blouet, R., Mokbel, C., Mokbel, H., Sanchez Soto, E.,

Chollet, G., & Greige, H. (2004). BECARS: a Free

Software for Speaker Verification. In Proceedings of

The Speaker and Language Recognition Workshop,

ODYSSEY, pp 145-148.

Coccagna, G., Pollini, A., & Provini, F. (2006).

Cardiovascular disorders and obstructive sleep apnea

syndrome. In Clinical and Experimental Hypertension

Vol. 28:217–24.

Do, M. N., (2003) Fast approximation of Kullback-Leibler

distance for dependence trees and Hidden Markov

Models. IEEE Signal Processing Letter 10, 115-118.

Fernandez R., Hernández L. A., López E., Alcázar J.,

Portillo G., & Toledano D. T. (2008). Design of a

Multimodal Database for Research on Automatic

Detection of Severe Apnoea Cases. In Proceedings of

6th Language Resources and Evaluation Conference.

LREC, Marrakech.

Fiz, J.A., Morera, J., Abad, J., Belsulnces, A., Haro, M.,

Fiz, J.I., Jane, R., Caminal, P., & Rodenstein, D.

(1993). Acoustic analysis of vowel emission in

obstructive sleep apnea. In Chest Journal; 104: 1093 –

1096.

Fox, A.W., & Monoson, P.K. (1993). Speech dysfunction

of obstructive sleep apnea. A discriminant analysis of

its descriptors. In Chest Journal; 96(3): 589-595.

Fredouille, C., Pouchoulin, G., Bonastre, J.F., Azzarello,

M., Giovanni, A., & Guio, A. (2005). Application of

Automatic Speaker Recognition techniques to

pathological voice assessment (dysphonia). In

Proceeding of 9th European Conference on Speech

Communication and Technology, Interspeech 2005,

Lisboa, p. 149-152.

Glass, J. , & Zue, V. (1985). Detection of nasalized

vowels in American English. In Proceedings of

Acoustics, Speech, and Signal Processing, IEEE

International Conference on ICASSP, Volume: 10, p.

1569- 1572.

Hidalgo, A., & Quilis, M. (2002). Fonética y fonología

españolas. Editorial Tirant blanch.

Lloberes, P., Levy, G., Descals, C., et al. (2000). Self-

reported sleepiness while driving as a risk factor for

traffic accidents in patients with obstructive sleep

apnoea syndrome and in non-apnoeic snorers. In

Respiratory Medicine 94: 971–6.

Moreno, A., Poch, D., Bonafonte, A., Lleida, E., Llisterri,

J., Mariño, J.B., & Naude, C. (1993). ALBAYZIN

Speech Database: Design of the Phonetic Corpus. In

Proceedings of Eurospecch 93. Berlin, Germany, 21-

23. Vol. 1 pp. 175-178.

Parsa, V., & Jamieson, D. G. (2001). Acoustic

discrimination of pathological voice : Sustained

vowels versus continuous speech. In Journal of

Speech, Language, and Hearing Research ISSN 1092-

4388, vol. 44, no2, pp. 327-339 (1 p.1/4).

Pruthi T. (2007) Analysis, vocal-tract modeling and

automatic detection of vowel nasalization. Doctor

Thesis at the University of Maryland.

Puertas, F.J., Pin, G., María, J.M., & Durán, J. (2005).

Documento de consenso Nacional sobre el síndrome

de Apneas-hipopneas del sueño (SAHS). Grupo

Español De Sueño (GES).

Reynolds, D.A., Quatieri, T.F., & Dunn, R.B. (2000).

Speaker verification using adapted gaussian mixture

models. In Digital Signal Processing 10: 19-41.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

130