ACOUSTIC AND PERCEPTUAL CHARACTERISTIC

OF ITALIAN STOP CONSONANTS

Loredana Cerrato, Mauro Falcone

Fondazione Ugo Bordoni
Speech Communications Group

ABSTRACT All the target syllables were inserted in stress position in real
words which were embedded in meaningful sentences such as: Il
We report in this paper the results of a study carried out to papà di ADA ha zAPPAto nell'orto con la zAPPA rotta.
analyse the acoustic and perceptual characteristics of Italian stop
consonants. The aim of this study is twofold: give an acoustical 2.1. Corpus for the acoustic analyses
description of Italian stops and investigate which are the
perceptual cues relative to their place of articulation. The material used for the preliminary acoustic analysis was
extracted from the production of VCV and VC:V sequences by
From the acoustic point of view we report: the measurements all the 10 speakers in two different speaking styles: semi-
relative to the length of the whole consonant and of its release spontaneous speech and read speech.
burst; the F1 and F2 of the following vowel measured at the
beginning of it. Moreover we counted the presence of the release The parameters we analysed are:
burst and we tried to describe its acoustical characteristics in 1. total duration of consonantal segment (interval
terms of the spectral structure as suggested by Blumstein [1] [2]. between the end of the preceding vowel and the
From the perceptual point of view we report the results of three beginning of the next vowel);
perceptual tests that we run with the aim of evaluating whether 2. burst length (if present);
the release burst or the formant transitions are more relevant for
the perception of Italian stop consonants’ place of articulation. 3. frequency of F1 and F2 of the following vowel,
measured at the beginning of the vowel;
1. INTRODUCTION
4. description of the burst in terms of its structure.
Although it is commonly agreed that the acoustic cues which
make the identification of stop consonants possible lie in the The results of the acoustic analysis in the time domain (points 1
burst portions and in the adjacent transition segments, there is no and 2) are reported in table 1 and 2.
unanimity as to the relative contribution of each cue [3] [4] [5]
[6]. In particular for Italian there are very few studies that CONSONANT DURATION spontaneous read
investigated the acoustic information of stop consonants [7] [8] Single stop 66 m 71 ms
[9], our study represents a pioneering work which gives some Geminate stop 134 ms 112 ms
insight into the relative importance of the different acoustic cues Table 1: Average values of the syllable duration of the Italian
of stop consonants. stop consonants measured in our corpus, for semi-spontaneous
and read speech.
2. SPEECH MATERIAL
BURST DURATION spontaneous read
We had to build an ad hoc corpus for our study as all the
Single stop 7 ms 8 ms
available databases didn't seem to contain enough material for
our scope. We created a kind of "building-sentences task" to Geminate stop 11 ms 10 ms
elicit semi-spontaneous speech from 10 speakers (5 male and 5 Table 2: Average values of the burst duration of the Italian stops
female university students from the area of Rome) who were measured in our corpus, for semi-spontaneous and read speech.
recorded in our labs while trying to build a series of 12 No other European languages, among the most common ones,
sentences containing target syllables. have geminated consonants like Italian. For this reason, no
Target syllables were VCV and VC:V, with V=a and C= /b d g p studies have been conducted to analyse and describe geminated
t k/ and C:=/b: d: g: p: t: k:/ that is the complete set of the Italian stop consonants. Our analysis show that geminated stops have a
single and geminated stop consonants. The VCV and VC:V total length which is almost the double of single stops, and the
sequences all have the same structure /a/+stop+/a/, in order to release burst appears to be longer in geminated stops showing at
minimise the acoustic and phonetic variability due to the times a particular “double” realization.
coarticulation phenomena. The main results relative to the frequency domain are reported in
table 3.
 VOWEL F1 (Hz) F2 (Hz) 3. The third set, which we call transitions, consist
reference /a/ 600 1500 of the same stimuli forming the syllable set,
pa 620 1145 from which, this time we took out the short
ba 620 1320 segment containing the release burst. The length
of this stimuli is about 225 ms.
ta 620 1420
da 620 1545
2.3 Editing criteria
ka 630 1610
ga 640 1585 The target words were excised from the sentences with the aid
Table 3: Average values of the first and second formant of the of a digital computer program, which displays the waveforms
following vowel, measured at the beginning of it. and spectrograms of the syllables to be analysed and plays them
if required.
The value of F2 represents the transition from the consonant to
the vowel, therefore it varies according to the place of Attention was paid to cut the waveform where the value is as
articulation of the preceding consonant. For instance in the case closest to zero as possible.
of bilabial stops /b, d/, that have a very low place of articulation,
as the F2 value is lowered towards their articulation point. Syllable set

The results relative to the spectral structure of the release burst From the target words we selected the syllables with the
are quite consistent with those reported in [1]: following criteria: the beginning of the stimulus was edited from
the end of the transitions with the previous sound and the end
 DA diffuse raising pattern with energy was marked at the beginning of the transition with the following
around 2.5 kHz for dental consonants; sound.

 DD diffuse falling pattern with energy Burst set

around 1.5 kHz for labials;
From the syllable stimuli, we extracted the burst stimuli with the
 C compact pattern with energy around 1.7 following criterion: we started to edit the signal portion from the
kHz for velars. end point, that is where on the waveform it is visible the
beginning of the final vowel of the VCV syllable and from a
We also counted the presence of the release burst in all the stop perceptual point of view it is hearable the sound of the vowel.
consonants. The results, reported in table 4, show that it ranges
from 66% to 96% depending on the style of speech (semi- Usually this point corresponds to the second periodic peak
spontaneous vs. read) and on the type of consonant (single vs. visible on the waveform, while the first peak should represents
geminated, bilabial vs. dental vs. velar). the burst. Starting from this point and going backward on the
time axis for about 25ms we selected our stimuli. This
Percent of burst presence spontaneous read segmentation assures that always the same part of the consonant-
Single stop 66 % 85 % vowel transition is included in the stimuli, that is only the first
Geminate stop 95 % 96 % two visible peaks on the waveform. Moreover when the burst is
not visible on the waveform this criterion it still allows to
Table 4: Percent value of the presence of the release burst.
accomplish the segmentation (see fig.1 and fig 2).
2.2 Corpus for the perceptual tests
From the speech material gathered for this first analysis, we
extracted a smaller corpus consisting of 6 VCV e 6 VC:V
syllabic segments containing the 12 Italian stop consonants
uttered in semi-spontaneous speech by 2 male and 2 female
speakers. In total we selected 48 stimuli. With this corpus we
created three sets of stimuli for three different perceptual tests.

1. The first set, which we call syllable, is made of

syllabic stimuli having a slightly variable
duration (±1ms) which is around 250 ms.

2. The second set, which we call burst, is made of

short stimuli, of the length of about 25 ms Figure 1: A typical segmentation of a CVC stimulus.
(±1ms) which we extracted from the syllable The highlighted portion in the waveform is the selected burst
stimuli. They consist of the short portion stimulus.
containing the release burst relative to the stop
consonants.
Of course, it could be argued that, in order to evaluate the
perceptual importance of the release burst we shouldn’t have
included any part of the previous or following transition. But deprived of the segment containing the release burst are
this would not be possible as the effective duration of the burst presented.
alone is on average 10 ms, which is practically too short to be
perceivable. In total, without considering the confusion between single and
double consonants (see column % correct+ in table 5), the
results of the syllable test show a correct perception of the
stimuli for 94%. If we consider also the mistakes due to the
confusion between single and double consonants (see column %
correct in table 5), we still have a high percentage of correct
answers 75%, and the 20% of incorrect answers are due to the
confusion between singles and geminates and vice versa.

In particular the results of the syllable test show that only two
types of mistakes occurred: /t/ perceived as /d/ with 1.46%; /k/
perceived as /g/ 2.71%. This confusion occurred mainly because
in semi-spontaneous speech unvoiced stop consonants tend to
have an incomplete closure, or no closure at all, sounding more
like fricatives than stops; but as we didn’t put the fricatives
Figure 2: Detail of the release burst waveform. among the possible answers, subjects identified these
fricativised unvoiced stops as the relative voiced stop.
Transition set
TEST TYPE % correct % correct+
These stimuli consist of the same stimuli forming the syllable Syllable set 74.4 94.2
set, from which we took out the short segment representing the
Transition set 51.1 76.8
release burst, as a consequence they have a length of 225 ms.
Burst set 12.8 25.3
Table 5: Summary of the results obtained in the three subjective
3. EXPERIMENTS SET UP
tests. The value % correct is the percent of stimuli correctly
Three perceptual test were run, one for each set of stimuli: perceived. The % correct + is the percent of percent of stimuli
correctly perceived allowing confusion between a single and its
1. test syllable geminate or vice-versa.
2. test transition 10 0

3. test burst
80
3 different groups of 20 listeners (university students, aged
between 21 and 30) served as subjects for each test. They had no 60
hearing pathologies and nobody was an expert in perceptual
phonetics. In all the tests they listened to the stimuli which were
40
presented in a random order.

Each stimulus was repeated three times with a short pause of 20

approximately 0.5 seconds and it was introduced by a voice
announcing it by a sequential number. At the end of each 0
stimulus presentation the listeners had to make a choice among b bb d dd g gg p pp t tt k kk

the twelve possible Italian stops (six singles and six geminates) sillab le 33 90 10 0 93 95 60 83 68 80 83 63 48

reported on an answer sheet. Before carrying out each test, the t r ansit io n 21 85 94 80 88 60 61 29 28 36 23 10

subjects underwent an initial phase of training of the duration of b ur st 18 11 14 17 9 11 25 19 3 10 8 9

approximately 5 minutes. They listened to the stimuli over

headphones, in binary modality with fixed volume at a Figure 3: Percent of correctly perceived stimuli in the three
comfortable level about 73dB SPL.
subjective tests for all the single and geminate stop consonants
of the Italian language.
4. RESULTS
The results of these tests, summarised in table 5, show that the
accuracy of identification of the consonantal place of In figure 3 it is reported a very high degree of confusion
articulation from burst stimuli only is very low (25%). The between the perception of /b/ and /b:/ in both syllable and
accuracy of identification increases dramatically when the transition stimuli. This is due to the fact that the speech material
subjects are presented with stimuli made of the burst and the we used is affected by a specific phenomenon of regional
transition with the following vocalic sound. Moreover the pronunciation: the very low correct perceived rate (33%) of /b/
accuracy of identification rises even when VCV syllables depends on the production of the voiced bilabial /b/ that in the
central Italian regional variation (and in particular in the area of 5 CONCLUSIONS
Rome) is usually produced as geminated.
Deriving conclusions from the results of our tests, we have to
The results for the transition test need further discussion. While underline that the limited speech material set we used might
the deletion of the burst affects only slightly the performance of have affected our results. Our corpus was constrained by three
listeners, who can still correctly perceive the voiced stop different factors: the fixed context of the stop consonant in the
consonants /b,d,g/, on the contrary the perception of the VCV syllable (V=a), the speech modality that is near to
unvoiced stops /p,t,k/ and their geminates, appears to be affected spontaneous speech and the limited speaker class (only young
by the deletion of the burst, with the consequent loss of about 10 speakers from the same area). For a wider investigation less
to 50 percent in correct perceived stimuli. It is important to constraints are of course necessary. Nevertheless the results of
outline that these errors are not equally distributed in the set of this study provide some interesting information on the acoustic
stimuli. There are in fact four stimuli which are always and perceptual cues of Italian stops, and in particular that: Italian
misperceived by all listeners: a /k/ perceived as /g/ (or /g:/), a stop consonants in syllabic context are correctly identified by
/k:/ perceived as /t/ (or /t:/), a /p/ perceived as /b/ (or (/b:/), and listeners, while the stimuli representing the release burst seem
finally a /t/ perceived as /d/. As all listeners perceive these not to be sufficient for the correct identification of the stop
stimuli in the same wrong manner, we believe that the stimuli consonants’ place of articulation. Moreover with the transition
really “sound as” a different stop consonant. In this case it is stimuli we obtain a very good percentage of correct
hard to consider this result as an error. Probably the identification, which support the hypothesis that the acoustic
manipulation of the speech signal, i.e. the removal of the burst, information relative to the place of articulation of Italian stop
deeply affected the nature of these stimuli. consonants doesn’t lie in the release burst portion, but, in the
transition with the previous and following vowel [6,8].
100
5. REFERENCES
80 1. Blumstein, S., Stevens, K., “Acoustic invariance in
speech production: evidence from measurements of
60 spectral characteristics of stop consonants”, JASA 66,
pp.1001-1017, 1979
40
2. Blumstein, S., Stevens, K. “Perceptual invariance and
onset spectra for stop consonants in different vowel
20 environments”, JASA. 67, pp.648-662, 1980

0
3. Cooper, F., Delattre, P.C., Liberman, A. M., Borst, J.
b+ d+ g+ p+ t+ k+ M., Gerstman, L., “Some experiments on the perception
sillable 100 98 100 93 91 84 of synthetic speech sounds”, JASA. 24, pp.597-606,
transition 99 95 99 77 45 45 1952
burst 25 30 19 40 16 21 4. Liberman, A.M.,“Some Results of Research on Speech
Perception”, JASA. 29, pp.117-123, 1957
Figure 4: Percent of correctly perceived stimuli, allowing
5. Bonneau, A., Djezzar Laprie, Y., “Perception of the
confusion between singles and geminates, in the three subjective place of articulation of French stop burst”, JASA. 100,
tests. pp.555-564, 1996
6. Kewley-Port, D., “Representation of spectral change as
The percentage of the correctly perceived stimuli in the three cues to place of articulation in stop consonants”,
tests is reported in figure 4. In the burst test the result is so low Technical Report n.3 Research on speech perception,
that there are no evidence to support any hypothesis. In fact if Bloomington Indiana University Press, 1980
we consider that a random choice will have a value around 8%, 7. Cerrato, L., Falcone, M., “Il burst nelle occlusive in
it is clear that the obtained value of 13% is not indicative of any sequenze VCV e VC:V dell’italiano: un’analisi
relation between the perceived stimulus and the subjects’ choice. acustica”, Atti delle VIII° Giornate di Studio del Gruppo
None of the single or geminate stops have a high score of correct di Fonetica Sperimentale (in press), Pisa 1997
perception. This means that the burst itself does not deliver any 8. Albano Leoni, F., Maturi, P., “Forma e sostanza nei
information about the stop consonant’s place of articulation suoni del linguaggio”, in L’Interfaccia tra fonetica e
independently of its phonetic characteristic. In other words the fonologia E. Magno Caldognetto (a cura di) Studi di
judgement given in this test appears to be almost a random Linguistica applicata Unipress, pp.115-126
choice.
9. Landi, R., “Le consonanti occlusive in stili differenti di
parlato”, Atti delle 7° Giornate di Studio del GFS 1996,
pp. 143-155, Napoli, 1996