RECORDING SPEECH SOUND AND ARTICULATION IN MRI

Daniel Aalto, Jarmo Malinen, Pertti Palo

Dept. of Mathematics and Systems Analysis, Aalto University, P.O. BOX 11100, FI-00076 Aalto, Finland

Olli Aaltonen, Martti Vainio

Institute of Behavioural Sciences, University for Helsinki, P.O. BOX 9, FI-00014 Helsingin yliopisto, Finland

Risto-Pekka Happonen, Riitta Parkkola, Jani Saunavaara

Dept. of Oral Diseases, Dept. of Radiology, Medical Imaging Center of Southwest Finland

University of Turku, FI-20014 Turun yliopisto, Finland

Keywords:

MRI, Sound recording, Speech, Vowel, Formant.

Abstract:

This article describes an arrangement for simultaneous recording of speech and the geometry of vocal tract.

Experimental design is considered from the phonetics point of view. The speech signal is recorded with an

acoustic-electrical arrangement and the vocal tract with MRI. Finally, data from pilot measurements on vowels

is presented, and its quality is discussed.

1 INTRODUCTION

Helmholtz (1863) put forward the acoustic theory of

vowels by showing that the perceived vowel quality

depends on the resonance characteristics of the vo-

cal tract. Since then, it has been the main approach

to the acoustic theory of speech production (see, e.g.,

(Fant, 1960)). Based on these ideas, Mrayati, Carr,

and Guerin (1988) presented the Distinctive Regions

Model (DRM) of speech production suggesting that

speech production derives from regions that closely

correspond to established vowel and consonant places

of articulation.

Based on these earlier ideas, models, and theories,

it is hypothesized that the vocal tract conﬁguration

can be estimated strictly on the basis of the formant

structure. Here we present a data acquisition frame-

work for a mathematical model that not only solves

the direct problem of simulating speech sound from

a given 3D vocal track conﬁguration, but also allows

the prediction of vocal tract shapes on the basis of res-

onances corresponding to vowel formants — the main

information bearing parameters in speech.

Such simulators (comprising only of the wave

equation in the vocal tract) have been used for

studying normal speech production acoustics (Han-

nukainen, Lukkari, Malinen, & Palo, 2007; Lu,

Nakai, & Suzuki, 1993;

Svancara, Hor´aˇcek, & Peˇsek,

2004). When soft tissue and muscle models are in-

corporated, we expect that such a simulator is use-

ful for studying speech production from a wider pho-

netics point of view, and planning and evaluating

oral and maxillofacial surgery (Dedouch, Hor´aˇcek,

Vampola, &

Cern´y, 2002; Nishimoto, Akagi, Kita-

mura, & Suzuki, 2004;

Svancara & Hor´aˇcek, 2006).

See also Vahatalo, Laaksonen, Tamminen, Aaltonen,

and Happonen (2005) and Niemi, Laaksonen, Pel-

tomaki, Kurimo, Aaltonen, and Happonen (2006) for

background.

A computational model of speech production

(such as discussed in Hannukainen et al. (2007), Aalto

(2009), and Aalto, Alku, and Malinen (2009)) must

be validated by comparing simulated sound to mea-

sured sound in some metric (such as the resonance

structure, i.e., the formants). Since the simulation is

based on anatomic data, the validation of the compu-

tational model depends on recording a coupled data

set: the speech sound and the precise anatomy which

produces it. This requires imaging the vocal and nasal

tracts from the lips and nostrils to the beginning of the

trachea. We chose to use magnetic resonance imaging

(MRI) technique because of its safety in contrast to X

168

Aalto D., Malinen J., Palo P., Aaltonen O., Vainio M., Happonen R., Parkkola R. and Saunavaara J..

RECORDING SPEECH SOUND AND ARTICULATION IN MRI.

DOI: 10.5220/0003137301680173

In Proceedings of the International Conference on Biomedical Electronics and Devices (BIODEVICES-2011), pages 168-173

ISBN: 978-989-8425-37-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

ray based CT-imaging.

Using MRI poses many restrictions. When a full

3D scan of the vocal tract is desired, the acquisition

time will necessarily be long. During this time, the

test subject needs to remain stationary; in particular,

all the parts of the speech apparatus have to remain

as stable as possible as well as the fundamental fre-

quency (f

). We discuss these matters in Sections 3

and 4. On the other hand, the test subject’s voice

is recorded simultaneously when his

vocal tract is

scanned using MRI. This sound recording is carried

out using an arrangement that has already been re-

ported by Lukkari, Malinen, and Palo (2007), and Ma-

linen and Palo (2009). For previous work on simi-

lar projects, see, for example,

Svancara et al. (2004),

Ericsdotter (2005), and Bresch, Nielsen, Nayak, and

Narayanan (2006).

The purpose of this paper is to outline experimen-

tal protocols for acquiring the above mentioned data

sets in high quality, discuss these protocols from a

wide range of perspectives, and ﬁnally present some

observations based on pilot data.

2 SPEECH RECORDING

2.1 Phonetic Materials

From an articulatory point of view, the main prob-

lem in acquiring simultaneous MRI and audio data

is the subject’s ability to alter their articulatory set-

tings very effectively in a manner that retains the main

acoustic characteristics of the speech sounds. It is

well known that human subjects adapt their motor be-

havior regarding both vision and audition (Houde &

Jordan, 1998; Kelso, Tuller, Vatikiotis-Bateson, &

Fowler, 1984). With regard to speech articulation, the

adaptation of motor behavior is particularly important

since the production goals (i.e., speech sounds) occur

in numerous different articulatory contexts. There-

fore, it can be expected that speakers are forced to

adapt their articulation even during sustained vowel

production as their articulatory setup changes due to,

e.g., the contraction of the thorax. This causes addi-

tional problems when both MRI and audio data are

acquired simultaneously and calls for careful design

of the speech materials.

There are two main sources for adaptation in sus-

tained vowel production: (1) the changing position of

the larynx due to changes in fundamental frequency,

We shall use a male subject in the pilot stage as analyz-

ing a male voice is generally easier than analyzing a female

voice.

cue signal MRI sequence

speech beginning speech end

Time (s)

14.298

Frequency

Figure 1: A spectrogram showing a full sound recording.

From left to right: cue signal for the subject (≈ 3.5 s from

the signal onset, overlapping speech for the last ≈ 500 ms);

the clean speech sample (≈ 500 ms); the speech and the

imaging noise (≈ 8 s); and the clean speech sample (≈

500 ms).

and (2) the changing shape of the vocal organs due

to the contracting thorax during a long exhalation.

The ﬁrst problem can be circumvented by having the

subject produce the sustained vowels with a station-

ary fundamental frequency (f

). The second problem

cannot be completelety avoided and it calls for other

means to control the matching of MRI and audio data.

A partial solution is to measure the articulatory move-

ments during the vowel production.

With the above problems in mind, we designed

a plan according to which the subject was asked to

produce the sustained vowels with two different sta-

tionary f

levels (110 and 137.5 Hz; corresponding to

notes A2 and C#3, respectively). Furthermore, two

separate MRI imaging techniques were used: (1) a

stationary 3D image of the vocal tract was produced;

and (2) a dynamic 2D image showing the sagittal sec-

tion of the vocal tract during the vowel production.

The two f

levels were used to study the effect of lar-

ynx position on the vocal tract shape whereas the dy-

namic 2D image sequence was used to study the effect

of the changing vocal tract shape due to the contract-

ing thorax.

2.2 Experimental Setting

Before the imaging sequence starts, the subject lies

inside the machine in supine position in the same way

as during a standard MRI procedure of the head and

neck area. In addition, the sound collector is placed

upon the MRI coil, and it is positioned in front of test

subject’s mouth. The subject is able to speak to the

control room through the sound collector all the time.

Moreover, he can hear instructions from the control

room as well as his own (de-noised) voice (with a de-

lay of ≈ 20 ms due to acoustic wave guides) through

earphones of the MRI machine.

Before the experiment, the subject is given a de-

RECORDING SPEECH SOUND AND ARTICULATION IN MRI

169

scription of what he is asked to do next. When the

subject indicates that he is ready the experiment is

started. First, the subject hears a sinusoidal cue signal

that gives him a count-down for starting the utterance

at the right time as well as the desired pitch, i.e., the

level of f

A typical sound sample, including the cue signal,

is represented in Figure 1. The MRI machine is op-

erated so that a 500 ms “pure sample” of stabilized

utterance is obtained immediately before and right af-

ter the MRI noise interval.

After each experiment, the image data is inspected

visually and the subject gives his comments. During

the whole imaging sequence the sound sample is lis-

tened by a trained phonetician in the control room,

and unsuccessful utterances are usually detected im-

mediately. Particular attention is paid to the phonation

type and nasality. When doubt arises, the formants of

the sample are extracted from the two “pure samples”

using Praat 4.6.15; see Figure 3 below.

2.3 Imaging Sequence

Experiments were performed on a Siemens Magne-

tom Avanto 1.5T scanner (Siemens Medical Solu-

tions, Erlangen, Germany). Maximum gradient ﬁeld

strength of the system is 33 mT/m (x,y,z directions)

and the maximum slew rate is 125 T/m/s.

12-element Head Matrix Coil was combined with

the 4-element Neck Matrix Coil in order to com-

pletely cover the anatomy of interest. Coil conﬁgura-

tion allowed the use of Generalized Auto-calibrating

Partially Parallel Acquisition (GRAPPA) technique to

accelerate acquisition. Technique was applied in all

the scans using acceleration factor 2.

3D VIBE (Volumetric Interpolated Breath-hold

Examination) was found out to be the most suitable

MRI sequence for the rapid 3D acquisition required in

this study. Basically, 3D VIBE is an ultra-fast gradi-

ent echo sequence with an isotropic resolution. In ad-

dition, the k-space scan is typically performed asym-

metrically in this sequence, which reduces the num-

ber of phase encoding steps in the slice-selection di-

rection leading to faster scan times. As the naming

of the sequence suggests, it was originally developed

for fast 3D imaging of the abdominal region where

breath-hold during the scan is essential. Sequence

parameters were optimized in order to minimize the

acquisition time. The following parameters allow

imaging with 1.8 mm isotropic voxels in just 7.6 s:

Time of repetition (TR) was 3.63 ms, echo time (TE)

1.19 ms, ﬂip angle (FA) 6

◦

, receiver bandwidth (BW)

600 Hz/pixel, FOV 230 mm, matrix 128x128, number

of slices 44 and the slab thickness of 79.2 mm.

When higher resolution is required, imaging with

1.2 mm isotropic resolution is possible in 17 s when

following changes to parameters are applied: TR

3.95 ms, TE 1.34 ms, matrix 192x192, and 64 slices.

Dynamic MRI scans were performed using seg-

mented ultrafast spoiled gradient echo sequence (Tur-

boFLASH) where TR and TE were minimized. This

sequence is typically used in cardiac studies but this

time magnetization preparation pulse was not applied.

Single sagittal plane was imaged with a pace of 5.5

images per second using parameters TR 178 ms, TE

1.4 ms, FA 6

◦

, BW 651 Hz/pixel, FOV 230 mm, ma-

trix 120x160, and slice thickness 10 mm.

2.4 Sound Recording

The MRI room presents a challenging environment

for sound recording. Use of metal components and

electronics is restricted inside the MRI room, and it is

completely excluded near the MRI machine for safety

and image quality reasons.

We use the sound recording arrangement detailed

in Lukkari et al. (2007), and Malinen and Palo (2009):

A two-channel sound collector samples the speech

and noise signals in a dipole conﬁguration. The sound

collector is an acoustically passive, non-microphonic

device which does not cause artifacts in the MR im-

ages, and it is also transparent in X-Ray CT. The

sound signals are coupled to a RF-shielded micro-

phone array by acoustic waveguides of length 3 m.

Again, the waveguides are acoustically passive and

linear, but their frequency response is far from ﬂat

because of the longitudinal resonances that, how-

ever, have been satisfactorily controlled by special

impedance terminations at both ends of the wave

guides. The microphone array inside its Faraday cage

lies at a safe distance from the main coil of the MRI

machine. The signals are coupled from the micro-

phone array to custom RF-proof ampliﬁer that is situ-

ated outside the MRI room. These analogue electron-

ics are used to optimally substract the noise channel

from the speech channel in real time.

The audio signal is digitized using a Digidesign

M-Box model 1, 24bit ADC, controlled by a Mac-

BookPro2,2 computer running MacOSX 10.4.9 and

Pro Tools LE 7.3.1.

2.5 Acoustic Noise in the MRI Room

Loud acoustic noise is familiar to anyone who has

undergone an MRI study. The noise originates from

the vibrations of the gradient coil support structure.

These vibrations are caused by the interactions be-

tween the pulsed magnetic ﬁelds created in gradient

BIODEVICES 2011 - International Conference on Biomedical Electronics and Devices

170

coils and the main magnetic ﬁeld.

The dipole sound collector and the analog noise

cancellation (as described in Section 2.4) takes care

of a good part of the noise at low frequencies, say

< 500 Hz. In particular, we are able to obtain sound

recordings of vowel utterances in real time, most of

which have positive S/N ratios during the MRI noise.

However, we are not always able to produce a signal

that would be directly (without further de-noising) us-

able for formant extraction by, e.g., the linear predic-

tion algorithm.

For DSP-based post-processing, we record a noise

sample from each MRI sequence and conﬁguration

used. During these recordings, the test subject lies

silently inside the MRI machine so that the acoustic

conditions are the same as in actual speech record-

ings. For comparison, some noise samples are also

collected by a directional microphone ≈ 3 m away

from the MRI machine. The frequency response of

the acoustic wave guides is measured in an anechoic

chamber. For further details, and for post-processing

and analysis of the data we refer to Aalto, Aaltonen,

Happonen, Malinen, Palo, Parkkola, Saunavaara, and

Vainio (2011).

The acoustic MRI noise is signiﬁcantly different

for different imaging sequences. Ultrafast sequences

— such as 3D VIBE used in this work — require max-

imal performance of gradient system both in terms of

slew rate and amplitude. This results in exceptionally

loud acoustic noise. We remark that even smallest

changes in parameters of a given MRI sequence may

change acoustic noise signiﬁcantly. It is thereby es-

sential to maintain sequence parameters and patient

positioning constant.

3 RESULTS

For successful data acquisition, it is necessary to iden-

tify relevant parameters concerning the whole exper-

imental setting that must be kept track of at all times.

It is not always a priori clear what should be regarded

as a relevant parameter, or what practical steps must

be taken to keep them under control.

Here, the relevant parameters can be divided into

two groups: (1) physiological parameters involving

the human subject; and (2) physical parameters of

the measurement equipment. As documented in Sec-

tions 2.2 – 2.5 above, we have spent much effort to

optimize and standardize physical parameters which,

indeed, we can control for, measure and compensate

to a considerable extent. Some physiological param-

eters have been considered in Section 2.1, and they

have a much more problematic nature. We proceed to

Figure 2: Mid-sagittal sections of an 8 s production of the

vowel [A] with a) f

= 110 Hz; and b) f

= 137.5 Hz. An

overlaid image is shown in c) to indicate their difference that

is visible as lighter gray. Notice in particular the difference

in position of lower lip, tongue blade and larynx.

discuss them in terms of our pilot experiments.

Let us start with the level of f

. The subject was

instructed to keep f

in a given reference value, and

he was able to do that with an error of ± 3 Hz in all

experiments. Figure 2 indicates that different levels

of f

result in visible differences of vocal tract con-

ﬁguration while uttering the vowel [A].

We remark that changes in sound pressure may re-

sult in a similar change as in Figure 2. To exclude

this, the subject tried to keep the sound pressure same

in both imaging sessions, but he did not receive any

feedback in that respect. However, the (subjective)

exhalation time was of the same length in both mea-

surements. After recording and compensating the fre-

quency response of the whole setup, the sound pres-

sure levelcan be extracted from recorded signals quite

precisely. It can be observed that the sound pressure

given by the subject always increases towards the end

of the sample.

The measured formants of [A] (corresponding to

the experiment giving Figure 2) are given in (F

, F

)

-plane in Figure 3. As explained in Section 2.2, we

obtain two “pure samples” of speech, and we there-

fore have two points (connected by a line) for each

experiment in Figure 3. We have used both 7.6 s and

17 s imaging sequences, two f

levels, and several

phonation types.

Let us next discuss the changing of shape of vo-

cal organs. Figure 3 gives us a ﬁrst indication that

the vocal tract geometry somewhat “creeps” for some

reasons during the 7.6 s MRI scan. We measured this

RECORDING SPEECH SOUND AND ARTICULATION IN MRI

171

600 650 700 750 800

1000

1050

1100

1150

1200

F1 / Hz

F2 / Hz

/a/ 137.5 Hz, 8s

/a/ 110 Hz, 8s

/a/ 110 Hz, 18s

/a/ 110 Hz, breathy, 8s

Figure 3: Formants of the vowel [A] produced with f

110 Hz and f

= 137.5 Hz based on 500 ms samples from

the beginning (marked with *) and the end (marked with o)

of the sustained phonation. During these samples the MRI

sequence was not active.

phenomenon for vowel [æ] directly using a dynamic

MRI sequence in the sagittal plane; see Figure 4.

Figure 4 indicates quite large changes in both the

mouth and pharyngeal cavities caused by the different

position of the larynx during a long exhalation; in Fig-

ure 2(c) there is in addition a (phantom) doubling in

the size of the vocal folds. That the larynx has moved,

is also supported by the movement of vocal folds that

can be observed from Figure 4. As explained in Sec-

tion 2.1, some movement is to be expected because of

the contracting thorax. A part of the observed changes

may, however,be due to changing aerodynamicforces

in the vocal tract that arise from variable air ﬂow dur-

ing the MRI scan.

We note that the formants are considered as purely

acoustic parameters of the vocal tract geometry, and

— as such — they do not depend directly on dynamic

variables such as the sound pressure and the air ﬂow

but, instead, through detectable changes in vocal tract

geometry including the time-dependent acoustic ter-

mination due to the vocal folds. It is not clear whether

one should (for physically motivated reasons) aim at

constant sound pressure or at constant air ﬂow in a

measurement leading to Figure 3. We remark, how-

ever, that observing the ﬂow inside the MRI machine

is probably very challenging.

4 CONCLUSIONS

We have described experimental protocols, MRI se-

quences, and a sound recording system that can be

used for simultaneous sound and anatomical data ac-

Figure 4: Overlaid image showing the difference between

the ﬁrst and last frame of the 8 s dynamic image sequence

of [æ] with f

= 137.5 Hz. Differences are visible as lighter

gray.

quisition of human speech. The results and experi-

ences of a pilot experiment on vowel formants and

the corresponding vocal tract geometries have been

reported. Such data sets are intended for parameter

estimation, ﬁne tuning, and validation of a mathemat-

ical model for speech production as discussed in Sec-

tion 1. However, these methods have a wide range of

applications in phonetics and medicine.

Phonetic Remarks and Observations

The MR imaging poses severe problems with regard

to both speech production by the test subject and

sound recording. The MRI requires long sustained

vowel production and — at the same time — it pro-

duces high levels of acoustic noise which masks the

speech sound.

The phonetic and articulatory problems stem from

the inability of a subject to maintain a stable vocal

tract shape long enough: there is a trade-off between

image quality in terms of resolution and speech pro-

duction “quality” in terms of articulatory stability.

The duration of an MRI scan (such as considered in

Section 2.3 above) is 7.6–17 s, and the sound record-

ing time is ≈ 2 s longer than that; see Figure 1. The

subject should be able to maintain constant position,

conﬁguration of the vocal organs, all sound charac-

teristics, and the type of phonation during the whole

period. According to our experience, this is a difﬁcult

requirement even for a healthy subject.

We designed a set of recording materials to ad-

dress these problems phonetically as well as possi-

ble. Our work indicates that the problems cannot be

circumvented altogether. There are, however, simple

BIODEVICES 2011 - International Conference on Biomedical Electronics and Devices

172

means to further improve articulatory stability during

recordings. We propose, at least, the following:

1. The test subject should be familiar with the MRI

noise as well as the cue signal so as to perform

optimally in the experimental situation.

2. The intensity of the cue signal should match the

MRI noise so that the initial voice production can

be maintained, i.e., possible sound intensity ﬂuc-

tuation should be avoided.

3. The MRI noise itself is periodic and interferes

with the voice f

when they are close. Hence, the

cue should be matched with the MRI noise fre-

quency proﬁle.

4. The voice sample f

should be standardized but in

a way that depends on the test subject.

5. The cue signal should be longer to allow the sub-

ject more time to inhale.

6. Externally triggered MRI sequences can be used

to introduce noiseless pauses.

All in all, the results from the current experiment are

encouraging. They clearly point to directions where

the setup reﬁnements and better understanding will

iteratively approach a useful solution to the whole

problem.

REFERENCES

Aalto, A. (2009). A low-order glottis model with

nonturbulent ﬂow and mechanically coupled acoustic

load. Master’s thesis, TKK, Helsinki. Available at

http://math.tkk.ﬁ/research/sysnum/.

Aalto, A., Alku, P., & Malinen, J. (2009). A LF-pulse from

a simple glottal ﬂow model. MAVEBA 2009 (pp. 199–

202). Florence, Italy.

Aalto, D., Aaltonen, O., Happonen, R., Malinen, J., Palo,

P., Saunavaara J., & Vainio, M. (2011). Recording

speech sound and articulation in MRI. Analysis and post-

processing of audio-spatial data. In preparation.

Branderud, P. (2008). Personal communication.

Bresch, E., Nielsen, J., Nayak, K., & Narayanan, S. (2006).

Synchronized and noise-robust audio recordings during

realtime magnetic resonance imaging scans (L). Jour-

nal of the Acoustical Society of America, 120(4), 1791 –

1794.

Dedouch, K., Hor´aˇcek, J., Vampola, T., &

Cern´y, L. (2002).

Finite element modelling of a male vocal tract with con-

sideration of cleft palate. Forum Acusticum. Sevilla,

Spain.

Ericsdotter, C. (2005). Articulatory-Acoustic Relationships

in Swedish Vowel Sounds. PhD thesis, Stockholm Uni-

versity, Stockholm, Sweden.

Fant, G. (1960). Acoustic Theory of Speech Production.

Mouton, The Hague.

Hannukainen, A., Lukkari, T., Malinen, J., & Palo, P.

(2007). Vowel formants from the wave equation. Jour-

nal of the Acoustical Society of America Express Letters,

122(1), EL1–EL7.

Helmholtz, H. L. F. (1863). Die Lehre von den Tonempﬁnd-

ungen als physiologische Grundlage fr dieTheorie der

Musik. Braunschweig: F. Vieweg.

Houde, J., & Jordan, M. (1998). Sensorimotor adaptation

in speech production. Science, 279(5354), 1213.

Kelso, J., Tuller, B., Vatikiotis-Bateson, E., & Fowler, C.

(1984). Functionally speciﬁc articulatory adaptation to

jaw perturbations during speech: Evidence for coordi-

native structures. Journal of Experimental Psychology,

10(6), 812–832.

Lu, C., Nakai, T., & Suzuki, H. (1993). Finite element sim-

ulation of sound transmission in vocal tract. J. Acoust.

Soc. Jpn. (E), 92, 2577 – 2585.

Lukkari, T., Malinen, J., & Palo, P. (2007). Recording

Speech During Magnetic Resonance Imaging. MAVEBA

2007 (pp. 163 – 166). Florence, Italy.

Malinen, J., & Palo, P. (2009). Recording speech during

MRI: Part II. MAVEBA 2009 (pp. 211–214). Florence,

Italy.

Mrayati, M., Carr, R., & Guerin, B. (1988). Distinctive

regions and modes: a new theory of speech production.

Speech Communication, (7), 257–286.

Niemi, M., Laaksonen, J., Peltomaki, T., Kurimo, J., Aalto-

nen, O., & Happonen, R. (2006). Acoustic comparison

of vowel sounds produced before and after orthognathic

surgery for mandibular advancement. Journal of Oral &

Maxillofacial Surgery, 64(6), 910–916.

Nishimoto, H., Akagi, M., Kitamura, T., & Suzuki, N.

(2004). Estimation of transfer function of vocal tract ex-

tracted from MRI data by FEM. The 18th International

Congress on Acoustics, Vol. II (pp. 1473 –1476). Kyoto,

Japan.

Vahatalo, K., Laaksonen, J., Tamminen, H., Aaltonen, O.,

& Happonen, R. (2005). Effects of genioglossal mus-

cle advancement on speech: an acoustic study of vowel

sounds. Otolaryngology - Head & Neck Surgery, 132(4),

636–640.

Svancara, P., & Hor´aˇcek, J. (2006). Numerical Modelling of

Effect of Tonsillectomy on Production of Czech Vowels.

Acta Acustica united with Acustica, 92, 681 – 688.

Svancara, P., Hor´aˇcek, J., & Peˇsek, L. (2004). Numerical

modelling of production of Czech Wovel /a/ based on FE

model of the vocal tract. Proceedings of International

Conference on Voice Physiology and Biomechanics.

RECORDING SPEECH SOUND AND ARTICULATION IN MRI

173