The effect of intonation on perception of Cantonese lexical tones.

Full text

Page 1

The Effect of Intonation on Perception of Cantonese Lexical
Tones
Joan K-Y Ma, Valter Ciocca & Tara L Whitehill
Division of Speech and Hearing Sciences
University of Hong Kong
joanma@hkusua.hku.hk

Abstract
The aim of this study was to investigate the perception of
intonation-induced changes in tone in Cantonese. Previous
studies have showed that the F0 level and contour of tones are
likely to be modified by different intonations. As perception
of tones depends primarily on the fundamental frequency (F0)
pattern, it is likely that these intonation-induced changes in
F0 will affect listeners’ perception. In the present study,
speech materials of two different intonations (question and
statement) with six tonal contrasts placed at two different
positions (initial and final) were presented to ten listeners.
The listeners were asked to identify the tone of the target
word embedded in the stimuli. Results showed that overall
tones were accurately perceived, except at the final position of
questions, where tones 33, 21, 23 and 22 were confused with
tone 25. The perceptual patterns of individual listeners were
also analysed, with differences in tone identification strategies
identified across listeners.
1. Introduction
Cantonese is a tone language in which a contrast in tone marks
a difference in lexical meaning. Cantonese tones are
characterized by tone level (high, mid and low) and tone
contour (level, rising and falling) [1]. There are six contrastive
tones in Cantonese: high-level (55), high-rising (25), mid-
level (33), low-falling (21), low-rising (23) and low-level (22)
[2]. Cantonese tones are of relative relations and the F0 levels
are relative, not absolute.
The use of intonation in a tone language could be
potentially confusing as fundamental frequency (F0) serves
two similar but separate functions [3]. The variation of F0 at
sentential level marks intonation while the variation within a
syllable marks tone. Studies in different languages have
disagreed in terms of how tones are modified by intonation
(e.g. [3][4][5]), but it is generally agreed that tone is likely to
be modified by intonation (e.g. [4][7]).
In Cantonese, there have been a few studies on the effect
of intonation on lexical production using acoustic analysis.
Fok-Chan [1] studied tones produced in neutral, anger,
empathetic and interrogatory manners, and she found that all
the tones produced in an interrogatory manner have a rising F0
contour. Vance [3] examined the effect of word position
within statements and found the existence of sentence-final
tone lowering in Cantonese. In investigating the effect of
questions and statements on tone production, Ma, Ciocca &
Whitehill [8] showed that both tone level and tone contour are
modified by sentence intonation. They found that questions
raise the F0 level of the tone at all positions, when compared
to statements. In addition, at the final position of statements,
all six tones showed a rising contour regardless of the inherent
contour. These studies showed that the F0 patterns of tones are
differentially affected intonation when placed at different
positions of the sentence.
F0 is the most important perceptual cue in tone
perception [1][3]. Fok-Chan [1] suggested that once the
structure of the tone is disturbed, perception would be affected.
Fok-Chan [1] found that tones produced in interrogatory
manners were likely to be confused with tone 25 and tone 23,
and explained the perceptual pattern with the acoustic changes
observed in tones produced in interrogatory manner. However,
in Fok-Chan’s study [1], the stimuli used were produced in
isolation form, which might be different from tones produced
within sentences, where the context might provide cues to
listeners to compensate for the intonation-induced F0 changes
in tone. Fok-Chan [1] also mentioned that tones produced in
isolation are difficult to perceive accurately. This is because
the context in which the tone is produced provides the
listeners’ cues for tone normalization, which are reported to be
important for tone perception [9][10]. Therefore, stimuli
presented in isolation might underestimate listeners’ ability to
perceive intonation-induced changes in a natural context. The
purpose of the present study was to examine how intonation-
induced changes in F0 pattern of tone are perceived when
presented within a natural sentence.
2. Method
2.1. Listeners
Ten females served as listeners (aged 18 to 19 years old).
They were first-year undergraduates in the Division of Speech
and Hearing Sciences, University of Hong Kong. They were
considered naïve listeners, as the experiment was carried out
within their first two months at university, during which they
received limited training in phonetics.
Cantonese was the first language for all the listeners. All
listeners passed a hearing screening (≤ 20 dBHL at 250, 500,
1000, 2000 and 4000 Hz). A tone perception screening was
carried out with all listeners, consisting of 24 trials (four trials
for each tone). Listeners had to correctly perceive at least
three out of four trials per tone in order to pass the screening.
They achieved an overall average of 99.58% accuracy.
2.2. Speech Materials
Speech materials were collected from two speakers (one male
and one female). Two contexts were designed, with the target
word in initial position (/X tsi22 hou25 l n21 sΕ35/ ‘X is difficult
to write’) or final position (/lei55 k
is X’).
33 tsi22 h i22 X/ ‘This word

Page 2

Three sets of target words were embedded in the above
contexts. Each set consisted of six words that differed only in
tone. The three sets of words were derived from the syllables
/si/, /ji/, /j u/. All target sentences could be read in two
different intonations – question or statement. With two
contexts, eighteen target words and two intonation patterns,
there were a total of 72 different stimuli from each speaker.
2.2. Procedures
The experiment was carried out in a sound-attenuated room
(IAC sound-proof booth), with the speech materials presented
to the listeners through a Sennheiser HD 545 headset,
connected to an Apple Macintosh G4 computer. A HyperCard
programme was used to run the experiment. For each trial, six
Chinese characters were presented on the screen. The carrier
was also presented at the top of the screen.
The stimuli were divided into two blocks, one for the
male speaker and one for the female speaker. Within each
block, there were a total of 144 trials, as each of the 72 stimuli
was repeated once. The order of presentation (male, female)
was counterbalanced across the ten listeners.
Before each session, the 18 Chinese characters were read
aloud to the listeners by the first author to ensure that they
were familiar with all the characters. This was necessary as the
same Chinese character may have more than one
pronunciation in different contexts in Cantonese. Within each
trial, the listener was asked to identify the Chinese character
that matched with the word they heard by clicking on the
button representing the character. Each stimulus was presented
once and the listener could opt to listen to it a second time by
clicking on the “repeat sound” button. Each block took about
half an hour to finish, and each session took about one hour.
3. Results
3.1. Perceptual Accuracy
Confusion matrices were compiled separately according to
position (initial and final) and intonation (question and
statement). Group confusion matrices were compiled by
summing confusion matrices across the ten listeners, and are
displayed in Tables 1 to 4. The number in the cells represents
the percentage of responses being realized as that particular
tone. For example, tone 25 at the initial position of questions
was accurately identified 116 times out of 120 trials, resulting
in an identification percentage of 96.67%. The percentages of
correct identification are shown by the numbers appearing on
the diagonal of the matrices (shaded grey).
Table 5 shows the mean and standard deviation (SD) of
the percentage accuracy of the six tones. A series of Wilcoxon
matched pair tests were used to compare the accuracy of the
six tones. The percentage accuracy of tone 55 was
significantly higher than the other five tones (T = 0, p < 0.05
for all). Tones 25, 33 and 21 showed similar percentages of
correct identification with no significant differences between
them (p > 0.05 for all). Tone 22 showed accuracy that was
significantly lower than tones 33 (T= 6.5, p < 0.05) and 21 (T
= 0, p < 0.05), but no significant difference was found
between tones 25 and 22 (T = 11, p > 0.05). Tone 23 was
perceived with accuracy significantly lower than all the other
five tones (p < 0.05 for all), except for the difference between
tones 23 and 22 where no significant difference was noted (T
= 6.5, p > 0.05).



Tables 1-4: Confusion matrices for the six tones at different
positions (initial and final) and intonations (questions and
statements). Target tones are on the vertical axis, and
perceived tones on the horizontal axis. Cell numbers represent
the percentage of responses for each target tone.

Table1: Initial position of questions.
Target
Perceived Tones
55 25
55
100
25

96.67
33

99.17
21

23
2.50
22
6.67

Table 2: Initial position of statements.
Target Perceived Tones
55 25
55
100
25

99.17
33

95.83
21

23
0.83
22


Table 3: Final position of questions.
Target Perceived Tones
55 25
55
100
25

78.33
33
30.83
65.83
21
70.83
23
75.001.67
22
60.837.50

Table 4: Final position of statements.
Target
Perceived Tones

55 25
55
100
25

97.50
33

94.17
21

23
2.50
22


Table 5: Mean and standard deviation of the percentage
accuracy of the six tones
Mean (% Accuracy)
Tone 55100
Tone 2592.92
Tone 3388.75
Tone 2181.88
Tone 23 76.25
Tone 22 78.96
33


21



100


23

3.33
0.83

97.50

22





93.33



33


21



100


23

0.83
1.67

99.17

22


2.50


100



33


21

1.67

29.17


23

7.50
3.33

10.83

22

12.50



23.33

33


21



98.33

0.83
23

2.50
0.83

97.50

22


5.00
1.67

99.17



Standard Deviation (SD)
0
9.78
15.42
35.15
43.62
37.20
3.2. Intonation and Position
The percentages of correct identification were relatively
similar in all the contexts, except for the final position of
questions, where the percentage of correct identification was
low. At the final position of questions, tone 55 was perceived
with 100% accuracy in this context. For tones 25 and 33,
although the accuracy was lower than in the other contexts, the
differences between tones at final position of question and

Page 3

Table 6: Error pattern of the six tones at various contexts. Numbers in the cells represent the percentage of tone level (TL) error and
tone contour (TC) error.

Tone 55 Tone 25 Tone 33

TL TC TL TC TL
Question Initial
11.76
Question Final
26.47 50.00
Statement Initial
2.94 5.56
Statement Final
8.82 11.11


other contexts were not significant (p > 0.05 for all), except
for the difference between the final position of questions and
the initial position of statements for tone 25, where the
difference was significant (T = 0, p < 0.05). For the other three
tones (tone 21, tone 23 and tone 22), tones at the final position
were the least accurately perceived and were significantly
different from the other three contexts (T = 0, p < 0.005 for
all).
Tone 21
TL





Tone 23
TL

78.95

2.63

Tone 22
TL
7.92
8.91



TC
1.85
75.93
3.70
1.85
TC

97.70

2.30

TC
2.63
14.91
0.88


TC

82.18

0.99

3.3. Error Patterns
Error patterns in identification were also analysed. Each
response was classified as either correct or incorrect according
to the target tone. Incorrect responses were classified into
either (1) tone level error: the stimulus was identified as a tone
of the same contour, but an incorrect level of tone was
perceived (for example, tone 55 → tone 22), or; (2) tone
contour error: the stimulus was perceived as a tone of a
different contour (for example, tone 55 → tone 25). The error
patterns for each tone were summarized in terms of types of
errors and are displayed in Table 6. The numbers in the cells
represent the percentage of a given error type. For example,
out of the 34 identification errors for tone 25, 9 tone level
errors involving tones from the final position of questions
were observed. This gives a percentage of 26.47% of errors (9
out of the 34 errors) in this category.
As described in the previous section, stimuli presented at
the final position of questions were least accurately perceived.
Most of the errors in this condition involved misperceiving the
target tone as tone 25. This accounted for 75.93% of the
identification errors for tone 33, 97.70% of the identification
errors for tone 21, 78.95% of the identification errors for tone
23 and 82.18% of the identification errors for tone 22.
Differences in error patterns were noted between tones.
For tone 25, most of the errors were also found at the final
position of questions (76.47%). Among these final position
errors, 26.47% were tone level errors where tone 25 was likely
to be misperceived as tone 23 and 50% were tone contour
errors in which tone 22 was involved. For tone 33, perceptual
errors involving targets at the final position of questions
accounted for 75.93% of all the errors. There was a small
proportion of tone level error (11.11%) involving tones at the
final position of statements, where the level of the tone was
mistaken as low level (tone 22). No other major perceptual
error was noted for tone 21 besides misperceiving tone 21 as
tone 25 at the final position of questions. For tone 23, besides
confusing the level of tones at the final position of questions,
14.91% of errors at the final position of questions involved
tone 23 being identified as tone 22, a tone contour error.
Similarly, at the final position of questions for tone 22,
82.18% of all the perception errors were tone contour errors
where tone 25 was the likely response. Tone level errors were
also noted: of all the perceptual errors, 8.91% at the final
position and 7.92% at the initial position of questions
involved perceiving tone 22 as tone 33.
3.4. Identification Pattern for Individual Listeners
Confusion matrices of individual listeners were examined. The
percentages of correct identification of each listener at the
final position of questions are summarized in Table 7.
Listeners could be classified into four groups according to
perceptual patterns:
Group A. Listeners in this group were able to identify
tones 55, 33, 21 and 22 at the final position of questions
correctly, while their perception of tones 25 and 23 were
lowered. Two listeners were this category (L1 and L2).
Group B. Listeners in this group were able to identify
tones 55 and 25 correctly, but most of the responses for tones
33, 21, 23 and 22 were tone 25. The identification pattern of
four listeners fitted this description (L3, L4, L5 and L6).
Group C. In addition to being able to perceive tones 55
and 25 correctly (as for group B), this group of listeners was
able to perceive tone 33 with 100% accuracy, while their
perceptual accuracies for tones 21, 23 and 22 were lowered.
Two listeners were classified into Group C (L7 and L8).
Group D. As in Group C, tones 55, 25 and 33 were
perceived with a high level of accuracy. Although
demonstrating relatively low accuracy, this group of listeners
was also able to perceive tones 21, 23 and 22 accurately in
some trials. This group comprised two listeners (L9 and L10).

Table 7: Identification pattern for individual listeners.
Numbers in the cells represent the accuracy for that particular
tone by the listener on the left.
Listeners 55 35
L1
10033.33 100
L2
100 16.67100
L3
100 100
L4
100100 8.33
L5
100100
L6
100100
L7
10083.33100
L8
100100 100
L9
10075 100
L10
10066.67 100
33 21 23 22
91.67
91.67
0
0
0
0
8.33
0
50
50
33.33
8.33
0
0
0
0
0
0
50
16.67
83.33
91.67
0
0
0
0
0
0
25
25
0
0
50
4. Discussion
4.1. Perceptual Accuracy and Error Patterns
Except at the final position of questions, the six tones showed
identification accuracy above 90% accuracy in the different
contexts. Tones at the final position of questions had low
percentages of correct identification. This is presumably
related to the intonation-induced changes in F0 pattern at the
final position of questions.
At both positions (initial and final) of statements and the
initial position of questions, the error patterns observed were
consistent with those reported in Fok-Chan [1]. Confusion
between tones 33 and 22 was noted. Fok-Chan [1], who
reported a similar finding, suggested that this confusion is

Page 4

because both tone 33 and tone 22 are of slightly falling
contour, with difference only in F0 level. Tone 55 was
confused with neither tone 33 nor tone 22, even though tone
55 was also of a slightly falling contour, owing to the
distinctiveness in F0 level. Tone 23 was also misperceived as
tone 22 occasionally. Fok-Chan [1] assumed that this is
because of the gentleness of the slope for tone 23; listeners
would have to perceive an actual increase in F0 contour before
deciding it is a rising tone. A different finding was noted for
tone 25 at the final position of questions, where some trials
were perceived as tone 22. When the acoustic pattern was
explored in detail, tone 25 in this context began at an F0 level
lower than that of tone 22. It is postulated that some listeners
might focus on tone level, while the changes in tone contour
were compensated for by the F0 changes expected in this
intonation pattern (questions).
At the final position of questions, perceptual errors were
noted especially for tones 33, 21, 23 and 22. Ma et al. [8]
showed that all six tones in this context have rising contours,
regardless of the inherent F0 contour. Tones 25, 21, 23 and 22
were found to be very close in both tone level and tone
contour, while the level of tone 55 and 33 were distinctive
from the rest. Fok-Chan [1] claimed that once the structure of
the tone is disturbed, the identity of the tone will be lost and
perception will be difficult. This accounted for the relatively
well-preserved identification accuracy for tones 55, 25 and 33
in this context, but not for tones 21, 23 and 22. Although the
contours for tones 55 and 33 clearly differentiate from
canonical F0 patterns, listeners were able to use the distinction
in tone levels for these two tones as perceptual cues. For tone
25, the F0 pattern at the final position of questions coincided
with the rising contour of its original contour; therefore, the
perception in this context was not particularly affected. For
tones 33, 21, 23 and 22, most of errors at the final position of
questions involved misperceiving the target tone as tone 25.
With the similarity in F0 pattern, tones 21, 23 and 22 were
confused with tone 25. Although tone 33 in this context had a
tone level slightly above that of tone 25, the rising contour of
tone 33 in this context confused the listeners, as tone 25 is the
only rising tone with this F0 level in Cantonese. On the other
hand, a small portion of tone 25 targets was perceived as tone
23 in this context. This is probably related the closeness of
tone level between tones 25 and 23 at the final position.
4.2. Differences between Listeners
Gandour [11], who investigated the perceptual dimensions of
Cantonese tones, suggested that Cantonese tones were
perceived by features rather than by tone unit as a whole, as
the six tones showed differences in confusion patterns. The
results of his study suggested that the direction of F0
movement was more influential for listener perception than
tone level. Although the results from the present study showed
more tone level errors than tone contour errors, except at the
final position of questions, the differences in identification
pattern among listeners at the final position of questions
suggested that tone contour was not necessarily the most
salient cue in tone perception.
For the listeners in Group A, tone level was more
influential than contour in perception. Although all six tones
at the final position of questions showed a rising F0 contour,
the two rising tones (25 and 23) were the least accurately
perceived among the six tones; tones 55, 33, 21 and 22 were
perceived more accurately. As listeners’ perception was not
affected by the intonation-induced changes in F0 contour, this
suggests that the initial frequency of the tone and the tone
levels were more salient perceptual features for this group of
listeners. The context in which the tone was produced in might

provide listeners cues to compensate for the intonation-
induced F0 changes in perception.
By contrast, listeners in Group B placed more emphasis
on tone contour than level. Most of the trials for tones 33, 21,
23 and 22 were perceived as tone 25. This showed that the
rising contour have more saliency for this group of listeners,
rather than the contrast in tone level.
For listeners in Group C and D, both tone level and tone
contour impacted on perception. In addition to tones 55 and
25, tone 33 was perceived with 100% accuracy in these two
groups. Although the contour of tone 33 was modified to a
rising contour by intonation, the contrast in tone level was a
significant cue for the listeners to distinguish tone 33 from the
rest. However, these listeners showed poor ability in
distinguishing the overlapping contours of tones 21, 23 and 22
from tones 25 by the level cues (as for Group A). Tone
contour appeared to be a more salient cue for listeners in this
group, resulting in confusion of tones 21, 23 and 22 as tone 25.
Listeners in Group D showed a slightly better ability in
identifying tones 21, 23 and 22 than listeners in Group C, with
less than 50% identification accuracy of these three tones at
the final position of questions.
5. Conclusions
This study showed that intonation-induced F0 changes
listener’s perception of lexical tones. Tones at the final
position of questions were less accurately perceived than tones
in other contexts. Errors were especially noted for tones 33, 21,
23 and 22, where confusion with tone 25 was common. The
results also showed that different features (tone level and tone
contour) have different influences on tone perception.
Variations in the importance of these cues were noted across
listeners.
6. References
[1] Fok-Chan, Y. Y. (1974). A perceptual study of tones in
Cantonese. Hong Kong: University of Hong Kong Press.
[2] Chao, Y.R. (1947). Cantonese primer. Greenwood Press:
New York
[3] Vance, T. J. (1976). An experimental investigation of
tone and intonation in Cantonese. Phonetica, 33, 368-
392.
[4] Chang, C.T. (1958) Tones and intonation in the Cheugtu
dialect. Phonetica, 2, 59-85.
[5] Lyovin, A.V. (1978). Review of tone and intonation in
Modern Chinese by M. K. Rumjancev. Journal of
Chinese Linguistics, 6, 120-168.
[6] Ho, A.T. (1977). Intonation variation in a Mandarin
sentence for three expressions: interrogative, exclamatory
and declarative. Phonetica, 34, 446-457.
[7] Fry, D. B. (1968). Manual of phonetics. North Holland:
Amsterdam.
[8] Ma, K-Y.J., Ciocca, V. & Whitehill, T.L. (2004). The
effects of intonation patterns on lexical tone production
in Cantonese. Proceedings of International Symposium
on Tonal Aspects of Language: Emphasis on Tone
Languages, Beijing, China.
[9] Leather, J. (1983). Speaker normalization in perception
of lexical tone. Journal of Phonetics, 11, 373 – 382.
[10] Wong, P.C.M. (1998) Effects of speaking fundamental
frequency on the normalization of Cantonese level tones.
Journal of Acoustical Society of America, 104, 1834.
[11] Gandour, J. (1981). Perceptual dimensions of tone:
evidence from Cantonese. Journal of Chinese Linguistics,
9, 20-36