The importance of articulatory and phonetic training in the production of L2

40 
 
3. THE EXPERIMENT 
 
3.1 Aims and hypothesis  
The aims of this is study is to find out if an articulatory (ultrasound - US) 
training is better than a perceptual training and if a short training may help 
Italian learners in reproducing the English contrast /ɑ-ʌ/. These aims are 
based on the fact that Italian learners find extremely difficult to perceive and 
produce the contrast, since, according to the mother tongue phonetic-
phonological system, Italian learners produce /ɑ/-L2 as /ɔ/-L1 and /ʌ/-L2, 
often, as /a/-L1 (S. D’Apolito, 2017). The English orthography represents a 
further difficulty. Indeed, in Italian language the orthography is relatively 
transparent and native Italians tend to follow the orthographic form and, as a 
consequence, to mispronounce English written words, which are 
phonologically opaque instead. However, despite the fact that both trainings 
may help learners to move from the most similar L1 sound toward a new 
vowel target position, the US articulatory training is expected to be more 
effective than the perceptual one since a direct feedback can help learners to 
understand how to position the tongue for the production of the L2 vowels, 
or at least to differentiate them adequately from L1 similar vowels. The US 
is also expected to allow subjects to better control tongue position reaching a 
more stable tongue position over time. 
 
3.2 Method 
3.2.1 Subjects 
Nine female subjects from Southern Apulia (Salento) participated in the 
experiment. All speakers are monolingual, who have never been in a foreign 
country for longer than a month. The subjects have started studying English
41 
 
as a foreign language at the mean age of 8,5 and were mainly exposed to L1-
accented English teachers. Participants were divided as follows: i) three 
experimental subjects performed the articulatory US training (ES-US) i.e., 
SPK1, SPK2, SPK3; ii) three experimental subjects performed the perceptual 
training only (ES-P), i.e., SPK4, SPK5, SPK6; and iii) three control subjects 
(CS) who did not receive any training, i.e., SPK7, SPK8, SPK9. One female 
native speaker of American English was recorded to collect native speaker 
US data and 5 American English natives were recorded to collect acoustic 
data for perceptual stimuli to be presented during the training. 
3.2.2 Training procedures 
A one-hour training session for each type of training was planned. Both 
experimental groups received information about orthography, that is the 
grapheme-to-sound correspondences of the American English contrast, and 
the phonetic differences between the non-native vowels /ɑ-ʌ/ as well as with 
respect to the closest native vowels /a-ɔ/. Graphic representations of F1 
(tongue height) and F2 (tongue backness) vowel space of the American 
English contrasts and of the Italian vowels were provided followed by verbal 
instructions about their articulation. Then the two experimental groups 
underwent to different trainings. The ES-US group only received an audio-
articulatory training by means of the US machine (Toshiba Aplio XV), which 
offers a real-time biofeedback of learner’s tongue position together with a 
real-time movie of the native speakers’ tongue, used as a visual model. Thus, 
the US training started showing the native speaker’s movies (with audio) of 
tongue contour during the production of /ɑ-ʌ/ in isolation, then during the 
production of CVC words and finally the videos of entire real sentences. 
When subjects affirmed to have detected the acoustic and articulatory 
differences between the target vowels and between them and the native 
vowels /a-ɔ/, they started practicing the production of the L2 sounds with the 
US probe under their own chin. This allowed the subjects to see their own
42 
 
tongue profile on the screen and to adjust their tongue according to native 
speaker’s movement.  
The ES-P group only received a perceptual training, that is an identification 
test according to HVPT procedure. The perceptual training was performed by 
a web application expressly created by the CRIL laboratory staff. During  the 
training, participants were presented via headphones with one auditory 
stimulus at the time. The stimuli were English CVC words (V=/ɑ-ʌ/), 
produced by 5 English speakers and presented 4 times each.  Participants were 
asked to correctly associate the vowel sound to one of two non-orthographic 
symbols displayed on the computer screen. Non-orthographic symbols were 
used to avoid inaccurate associations between sounds and English 
orthography. A total of 150 trials was used in the training session (2 target 
vowels x 5 talkers x 3 contexts x 4 repetitions + 2 control vowels x 3 talkers 
x 3 contexts). 
3.2.3 Speech recordings 
Each recording session consisted in: i) pre-test data collection of L1 and L2 
production, ii) one-hour training (for ES-US, ES-P) and iii) post-test data 
collection of L1 and L2 production. The L1 corpus consisted of /pV1pV2/ 
words and pseudowords, where V1 was one of the five Italian native vowels 
/i, ɛ, a, ɔ, u/ and V2 was /i/ or /a/, proposed in a carrier sentence (e.g. “Dicevi 
pV1pi in su” or “Diceva pV1pa a Ken”). The L2 corpus consisted of /pV1p/ 
American English words, where V1 was /ɑ/ or /ʌ/, inserted in a carrier 
sentence (e.g. “I see pV1p inside”). All subjects read both corpora, displayed 
on a PC screen, 12 times. L1 and L2 productions of the three groups were 
collected in pre- and post-test sessions. Both data were collected 
simultaneously at CRIL laboratory in a soundproof room. US data were 
collected by the means of a convex probe positioned under the subject’s chin 
on the midsagittal plane, with a special stabilization set to fix the probe.
43 
 
Acoustic data were analyzed using PRAAT and the first three formants were 
calculated at the central 40% of the entire vowel duration. 
3.3 Acoustic data  
Acoustic data (22050Hz, 16 bit) were labelled and analyzed using PRAAT, a 
freeware program for the analysis and manipulation of acoustic speech signals. 
The program can generate spectrograms, intensity contour and pitch tracks, 
edit a recorded sound, label and extract individual sounds for further analysis, 
as shown in the lower figure. 
Figure 1: PRAAT’s graphical user interface 
 
The top pane is the waveform, which shows time left-to-right, and the local 
value of the sound signal in the up-and-down dimension. The middle pane is 
the spectrogram and the bottom pane is where the transcriptions are put . 
These options were helpful in labelling and calculating the first three formants 
(F1, F2, and F3), which were measured as the average of the values in the
44 
 
central 40% of the entire vowel duration. Successively, F1 and F2 mean and 
standard deviation values of the pre- and post-test target vowels were plotted 
on a Cartesian F1-F2 plan for each subject. Data were statistically analyzed 
by means of a series of independent test (p<0.05) in order to compare: 1) /ɑ-
ʌ/ in pre- vs post-test respectively; 2) /ɑ-ʌ/ in pre- and post-test vs L1 /a-o/ 
respectively.  
3.4 Articulatory data 
3.4.1. EdgeTrak 
The articulatory data were labelled using EdgeTrak, which is a computer 
program that automates the tracking of tongue contours by extracting (x, y) 
coordinates from the lower edge of the white curve in the ultrasound image. 
First, a few points on the first tongue frame of the sequence are manually 
chosen (fig. 2), and then EdgeTrak uses an active contour model to determine 
the location of the tongue edge in the image (fig.3). If the automatic tracking 
of the tongue edge does not produce satisfactory results, points can be 
manually added or subtracted to obtain the best fit. 
Figure 2: EdgeTrak 
manual contouring input. 
(ES-SPK4)
45 
 
  
Each following frame is determined by the same optimization process on the 
basis of the contour of the first one (fig.4). Automatic tracking accuracy is 
quite high, as the error between automatic and human expert tracking ranges 
from 1.83 to 3.59 pixels (1 pixel=0.295mm).  
 
3.4.2 Smoothing spline ANOVA and Bayesian intervals 
Subsequently, the frames with tongue surface contours were compared by 
exploiting the smoothing spline analysis of variance: the SS ANOVA. 
Smoothing splines are a type of natural cubic spline, which is a piecewise 
polynomial function that connects discrete data points called knots. The 
smoothing spline contains two terms, one that attempts to fit the data and one 
that penalizes a fit which does not have the appropriate amount of smoothness. 
Figure 3 (left) & 4 (right): EdgeTrak automatic tracking of the tongue edge, for the first 
and subsequent frames. (ES-SPK4)
46 
 
Although the penalty term does not allow the function to fit the data precisely, 
it ensures that the resulting spline has a suitable amount of smoothness (fig.5). 
The SS ANOVA is usually implemented in programming language and free 
software environment for statistical computing and graphics, such as R or S-
PLUS. Because of its benefits, the SS ANOVA statistical method has been 
used, as in this study, in applications that require a statistical technique to 
determine whether the shapes of multiple curves are significantly different 
from one another, taking into account their shape, rotation and translation and 
comparing each other as a whole.  
Figure 5: Smoothing spline estimate from twelve post-test repetition for L2 /ɑ/ (ES-
SPK4). The x axis is the length of the tongue, and the y axis is the height of the tongue. 
Like the ultrasound images, the tongue tip is on the right and the tongue root is on the 
left. 
 
Since the interaction may be significant even when a small section of the 
curves differs, Bayesian confidence intervals are used to determine whether 
the curves are significantly different at any point in the comparison. Given 
two groups of tongue contours, the SS for each contour set is termed main 
group effect, and around each SS the 95% Bayesian confidence interval is 
constructed. The comparison of the two groups of tongue curves is performed 
with the interaction diagram which represents a plot of the difference of the