PROSODIC ASPECTS OF CHILD-DIRECTED SPEECH IN CANTONESE

Introduction
The differences between the speech of an adult directed towards a child and the same adult's speech directed towards another adult have been investigated from many points of view - lexical, syntactic, phonological and phonetic to name a few. This paper presents the results of a study of some prosodic aspects of child-directed speech (CDS) in a language which hitherto has not been investigated, namely Cantonese.

The initial impetus for studies into CDS occurred in the 1960's as a reaction to innatist theories of language acquisition. Most of the work done in these early studies concentrated on English. The reader is referred to an earlier paper by Ogle and Maidment (1993) which includes a brief review of this literature and provides a summary of the main findings on CDS prosody. Languages other than English have also been the subject of study. Arabic and Spanish were investigated by Ferguson (1964), Latvian by Ru±e DraviÃa (1977), Marathi by Kelkar (1964) and German by Fernald and Simon (1984). A cross-linguistic study of CDS prosody including French, German, Italian, Japanese, British English and American English may be found in Fernald et al (1989). These and other studies broadly confirmed the results of the English studies. The salient characteristics of the prosody of CDS in these languages are: (a) an increase in mean fundamental frequency and (b) an expansion of the range of fundamental frequency which was effected by an increase in the upper limit of the range rather than a decrease in the lower limit.

That these tendencies are not perfectly universal is demonstrated by the results of the study of Quiche by Pye and Ratner (1984) which show that for this language mothers tend to lower mean fundamental frequency when speaking to their offspring. The study by Fernald et al (1989) showed that Japanese women tend not to expand their frequency range when speaking to their children. Even within one language it is clear that speakers vary considerably as to the extent of the upward shift of mean fundamental frequency. Shute and Wheldall (1989) report that some of their British English mothers showed little or no increase in mean or mode fundamental frequency for CDS while others showed marked increases.

The present study uses essentially the same technique and design as those employed by Ogle and Maidment (1993). This paper reported on fundamental frequency statistics for a small group of English-speaking mothers under two conditions: (1) spontaneous speech to another adult (henceforth the AA condition) and (2) spontaneous speech to a child (henceforth AC). The technique used to record fundamental frequency was electrolaryngography. The laryngograph is a non-invasive device which monitors vocal fold vibration by electrodes placed externally on the subject's neck at the level of the thyroid cartilage. The high frequency current which is passed through the larynx allows the measurement of the time-varying impedance of the neck caused by the vibration of the vocal folds. The resulting signal, known as Lx, can be recorded using a normal tape-recorder and processed further to derive an estimate of the duration of each glottal cycle (Tx) and instantaneous fundamental frequency (Fx). For further details of the operation of the laryngograph the reader is referred to Abberton and Fourcin (1984) and Abberton, Fourcin and Howard (1989).

One of the great advantages of laryngographic analysis is that it allows the rapid collection and analysis of a large number of data points. Whereas earlier studies were limited by their labour-intensive methods of analysis to basing statistical measures on quite small samples, Ogle and Maidment analysed samples of 4000-5000 points per speaker. The present study was able to use samples of approximately 8000 points.

In this paper results of the same kind as those presented for English by Ogle and Maidment are reported for Cantonese, namely mean fundamental frequency and range of fundamental frequency variation. The results obtained for the AA condition are compared with those for the AC condition. In addition, two further aspects of the prosody of Cantonese are investigated. These are (a): the difference in speech rate for the AA and AC conditions and (b) the differing realisation of lexical tone in the two conditions.

The first of these aspects has been investigated in the literature both qualitatively and quantitatively. Grewel (1959) characterised speech to the young child as slower in tempo with prolonged pauses between words, word groups and particularly between sentences. Sachs et al (1976) and Broen (1972) both report slower speech for CDS in terms of number of words per minute. Philips (1970) found that mothers took longer when reading particular sentences from a story to children than they did when reading the same sentences to another adult. As far as is known, there have been no investigations of the second aspect, lexical tone realisation.

Cantonese
Cantonese is a Yue language of Southern China and is spoken in Hong Kong, Guangzhou, Macau and in many overseas Chinese communities in Malaysia, Indonesia, USA and Britain. Like many of the languages of China, Cantonese is characterised by a simple syllable structure and by lexical tone. Lexical items which are segmentally identical may be distinguished by the use of differing pitch patterns. Cantonese has a particularly rich tone system. Most analysts agree that 6 distinct tones should be recognised, although the numbering and nomenclature of the tones vary somewhat from author to author:

Tone 1: high level

Tone 2: high rising

Tone 3: mid level

Tone 4: low falling (or extra low level)

Tone 5: low rising

Tone 6: low level.

The single syllable [fan] serves as an example of the lexical distinctiveness of these tonal patterns:

Further information on the structure of Cantonese may be found in Ramsey (1987) and Norman (1988).

Subjects and Recordings
Recordings were made of 7 Cantonese-speaking mothers, living in England. All the subjects were the carers of children aged between 1;0 and 1;8 at the time of the recording. The details are set out in Table 1 below.

Subject	Age of child	Sex of child
A	1;6	F
B	1;0	M
C	1;2	M
D	1;8	M
E	1;8	F
F	1;6	F
G	1;7	F

Table 1: Subject Details

The age of the children was chosen to be between 1 and 2 years as it was felt that features typical of CDS might have begun to disappear with older children. All of the speakers had lived in England for at least 3 years.

All recordings were made in the subjects' own homes so that they and their children would feel at ease. A portable laryngograph and Uher portable two-channel stereo tape-recorder were used. The speech signal was recorded using a clip-on microphone.

Two recordings were made, each of which lasted approximately 15 minutes:

1. The AA condition: the subject was recorded in conversation with the first-named author. The conversation concerned the subject's background, children and life in England.

2. The AC condition: the subject was recorded interacting with the child during story-telling. Three picture books were provided as a stimulus for the interaction. In some cases the interaction involved play with toys too.

Fundamental Frequency Statistics
Analysis and Results

A suite of programs, PCLX, running on a PC was used to analyse the Lx recordings. The Lx representation of larynx activity is first converted to a file of period measurements, Tx. The software provides facilities for analysing the Tx file in various ways. The most important analyses for the present purpose are those which result in a probability-density function of fundamental-frequency. The particular analysis used for this study gives a doublet distribution. Time-adjacent pairs of Fx values, derived from Tx, are examined to discover if they fall within the same subrange (frequency bin) of the range of fundamental frequency between 30 Hz and 1 kHz. There are 128 bins in this range and the centre frequency of the nth bin is given by:

The doublet distribution is used because it removes irregularity and rogue data points due to gross movements of the larynx and only takes into account portions of Fx which are constant or which display smooth and relatively slow change.

The PCLX software also provides statistics of the probability-density function of Fx. The two measures used in this study are the arithmetic mean and the 80% range. This latter is defined as the range between a lower limit which is centre frequency of the bin which has 90% of the samples above it and an upper limit which is the centre frequency of the bin which has 10% of the samples above it.

Table 2 below presents a comparison of the mean fundamental frequency in Hz of the 7 speakers under the two speaking conditions together with the percentage change from one condition to the other which is defined as ((AC-AA) * 100) / AA.

Subject	AA	AC	% change
A	244.0	296.0	21.3
B	213.0	225.0	5.6
C	213.0	265.0	24.4
D	225.0	237.0	5.3
E	191.0	219.0	14.7
F	191.0	196.0	2.6
G	213.0	219.0	2.8
Mean			11
Standard Deviation			8.4

Table 2: Mean fundamental frequency (Hz)

It is clear from Table 2 that all subjects used a higher mean fundamental frequency in CDS. The average increase was 11%. However, the variation between subjects was quite large. A Student's t test applied to the AA and AC figures yielded a significant result ( t = 3.06, p < 0.05).

It is of interest to compare the mean increase observed for these Cantonese speakers with the figure of 18.8% obtained by Ogle and Maidment for the English speakers in their study. The tendency for higher mean fundamental frequency is evidently considerably stronger amongst Ogle and Maidment's subjects than it is for the speakers of the present study.

Table 3 shows a comparison of the 80% fundamental frequency range in Hz for the 8 subjects in the AA and AC conditions. Again the percentage change from one condition to the other is presented.

Subject	AA	AC	% change
A	120.0	162.0	35.0
B	94.0	113.0	20.2
C	108.0	129.0	19.4
D	79.0	97.0	22.8
E	85.0	108.0	27.1
F	78.0	92.0	18.0
G	86.0	94.0	9.3
Mean			21.7
Standard Deviation			7.4

Table 3: 80% fundamental frequency range

As Table 3 shows, there can be no doubt that these speakers consistently used a wider range of fundamental frequency variation in CDS. Apart from speakers A and H, the size of the increase is quite consistent. An average increase in range of 27.1 percent was observed. A Student's t test on the data in Table 3 showed that the difference between the two conditions was highly significant (t = 5.17, p < 0.01).

Compared to the figure of 89.3% for Ogle and Maidment's English speakers, the average expansion of the 80% range is quite small for the these Cantonese speakers.

Table 4 presents a comparison of the Hz shifts in the upper and lower limits of the 80% fundamental frequency range for the AA and AC conditions. Once again the percentage change in frequency is given.

From the three lefthand columns in Table 4 it is clear that the expansion of the 80% range in CDS receives no consistent significant contribution from a downward shift in the lower limit of the range. In fact, all the subjects showed either no shift at all or an upward shift for the lower limit.

	Lower Limit			Upper Limit
Subject	AA	AC	%change	AA	AC	%change
A	201.0	237.0	17.9	321.0	399.0	24.3
B	186.0	191.0	2.7	280.0	304.0	8.6
C	180.0	219.0	21.7	288.0	348.0	20.8
D	201.0	207.0	3.0	280.0	304.0	8.6
E	166.0	180.0	8.4	251.0	288.0	14.7
F	166.0	166.0	0.0	244.0	258.0	5.7
G	186.0	186.0	0.0	272.0	280.0	2.9
Mean			7.7			12.2
Standard Deviation			8.2			7.4

Table 4: Comparison of upper and lower 80% range limits

The upper limit of the 80% range is shifted upward in CDS by all speakers. The mean figure for this shift is 12.1%, but there is considerable inter-speaker variation. The shift in the upper limit is shown by a t-test to be statistically significant (t = 3.64, p < 0.05).

These results are, in the main, consistent with those obtained by Ogle and Maidment, except that the shift in the upper limit is greater in most cases, as would be expected because of the very much greater expansion of the range displayed by the English speakers.

Speech Rate
There are many ways in which one could estimate the rate of speech. For the purposes of this paper an estimate of the number of syllables uttered per minute was chosen, mainly because it is a relatively simple measure to calculate for the fairly large amounts of speech involved in the recordings. Such a measure of course ignores possible variations of speech tempo and other aspects of potential interest such as pause length, occurrence of disfluencies and turn-taking behaviour. Nevertheless, any overall tendency of speech rate in CDS to differ from adult-directed speech rate should be captured by this measure.

The number of syllables in each sample was counted by repeated listening to the speech recording and this figure was divided by the total time for the recording. Table 5 shows the results for the two conditions and presents the percentage change between them.

Subject	AA	AC	% change
A	203	111	-45.3
B	189	110	-41.8
C	184	99	­46.2
D	174	101	­42.0
E	231	168	­27.3
F	198	156	­21.2
G	221	125	­43.4
Mean			­38.1
Standard Deviation			9.1

Table 5: Speech rate in syllables per minute

All speakers display a considerably slower rate for CDS. The difference between the two conditions is highly significant (t = -10.76, p < 0.01)

Tone Realisation
In order to assess the possible influences of CDS on the realisation of the 6 lexical tone patterns of Cantonese, two simplifying assumptions were made. The first of these is that the shape of all the tonal patterns can be adequately represented by their beginning and ending frequencies alone. The second assumption is that duration and intensity may be ignored in distinguishing one tone from another. Furthermore, in order to compare the realisation of tones across speakers, it is necessary to perform some normalising transformation of the frequencies in Hz. This transformation may be expressed as.

The start and end frequencies of the tones were estimated for each speaker under the two conditions by averaging over five tokens of each of the six tones. The boundaries of the relevant syllables were identified from a spectrographic display with playback facility. The system used in this part of the study did not allow simultaneous display of speech waveform and fundamental frequency. Therefore, the relevant frequencies were calculated spectrographically. Because of the possible interference of environmental noise on the speech recordings, the frequencies were estimated by measuring the tenth harmonic of a time-aligned narrow-band spectrogram of the Lx signal. Table 6 shows the start and end points of the six tones by speaker in the two conditions.

	Tone 1					Tone 4
	AAS	AAE	ACS	ACE		AAS	AAE	ACS	ACE
A	0.16	0.03	0.22	0.04		0.01	­0.23	­0.05	­0.22
B	0.30	0.12	0.37	0.15		-0.11	­0.21	­0.04	­0.20
C	0.28	0.18	0.26	0.18		-0.09	­0.22	­0.01	­0.25
D	0.25	0.14	0.28	0.09		-0.10	­0.27	­0.06	­0.21
E	0.32	0.17	0.25	0.14		-0.04	­0.17	­0.08	­0.21
F	0.22	0.07	0.26	­0.02		-0.10	­0.20	­0.07	­0.15
G	0.25	0.09	0.31	0.06		-0.14	­0.27	0.01	­0.24
Mean	0.25	0.11	0.29	0.10		-0.08	-0.22	-0.04	-0.21

	Tone 2					Tone 5
A	­0.16	0.04	­0.07	0.20		­0.10	0.01	­0.10	0.07
B	­0.06	0.17	0.02	0.27		­0.04	0.15	0.03	0.17
C	­0.08	0.14	­0.03	0.18		­0.05	0.08	0.00	0.12
D	­0.08	0.04	­0.06	0.17		­0.11	­0.04	0.02	0.11
E	0.02	0.27	­0.04	0.21		0.04	0.24	0.06	0.17
F	­0.07	0.13	­0.04	0.21		­0.05	0.04	­0.01	0.19
G	­0.03	0.15	0.07	0.36		­0.07	­0.02	0.10	0.21
Mean	-0.07	0.13	-0.02	0.23		-0.05	0.08	0.01	0.15

	Tone 3					Tone 6
A	­0.03	­0.11	0.06	­0.08		­0.02	­0.14	­0.08	­0.22
B	0.08	­0.03	0.10	­0.05		0.03	­0.08	0.07	­0.13
C	0.03	­0.04	0.13	0.04		­0.01	­0.08	0.02	­0.22
D	0.03	­0.15	0.11	­0.04		­0.01	­0.17	­0.01	­0.13
E	0.24	­0.02	0.10	­0.06		0.09	­0.04	0.09	­0.14
F	0.07	­0.03	0.14	­0.03		­0.03	­0.10	0.06	­0.11
G	0.06	­0.10	0.17	0.01		­0.02	­0.14	0.11	0.16
Mean	0.07	-0.07	0.12	-0.03		0.00	-0.11	0.04	-0.16

Table 6: Normalised frequency values for the six tones by speaker
(AAS/AAE = Adult-Adult start/end, ACS/ACE = Adult-Child start/end)

The above data is presented graphically in the appendix to this paper.

The following discussion is based upon Table 6.

The first point of interest is that three tones traditionally described as level (tones 1, 3 and 6) in fact show quite marked falling movements in both speech conditions.

It is noticeable that for all tones except tone 6 (low level), the AC start and end points are higher in normalised frequency than the corresponding AA start and end points. For tone 6, the start point is slightly higher for the AC condition, but the end point for AC is somewhat lower than for AA. This indicates that the speakers not only raise the overall fundamental frequency of their speech in CDS in absolute terms, but also adjust upwards the frequency of most individual tonal realisations within their fundamental frequency range. This adjustment is most marked for the two rising tones, tones 2 and 5.

The fact that in AC tone 1 and tone 3 (high level and mid level respectively) are adjusted upwards, considerably in the case of tone 3 and slightly in the case of tone 1, while there is no corresponding overall upward adjustment for tone 6 (low level), means that the space occupied by the subsystem of level tones is expanded in CDS and the distinctions between these three level tones are exaggerated.

This expansion of the level tone subsystem is evident from the averaged patterns in Figures 7 and 8 and is even clearer in some of the results for the individual speakers. Compare, for instance, the AA and AC conditions for subjects C and D.

Another readily identifiable difference between the two speech conditions concerns tone 2 (high rising). For all speakers this tone has the highest end point of the six tones in the AC. This is not always true for the AA condition. Moreover, there is a tendency for at least some of the speakers to increase the difference between the two rising tones (tones 2 and 5) in the AC condition. See especially subjects A, B and C.

Despite the above differences between the realisation of tonal patterns, it is remarkable that the overall relationships in the system (what Cheung, 1986 calls cardinal relations) are maintained. The following hold for both the AA and AC conditions:

(1) The relationship between the start frequency of the so-called level tones is Tone 1 > Tone 3 > Tone 6.

(2) The end frequency of Tone 4 is lower than that for Tone 6.

(3) The end frequency for Tone 2 is higher than that for Tone 5.

(4) Tone 2 has a steeper gradient than Tone 5.

(5) Tones 1 3 4 and 6 show a falling contour.

(6) Tones 2 and 5 show a rising contour.

The question arises as to whether the differences between the AA and AC realisations of tones show signs of lessening with the increasing age of the child. In order to test this possibility, the correlation co-efficient was computed between the age of the child in months and the absolute difference between the normalised frequency of the mid point of each AC tone and its corresponding AA tone. If the differences were age-sensitive, one would have expected a significant negative correlation. In fact, none of the correlation co-efficients was statistically significant. Therefore, one must conclude that over the age range investigated features of CDS affecting the realisation of tones are fairly stable and do not begin to abate.

In summary, it appears that the subjects in this experiment maintain the overall relationships of the adult tonal system in their speech to children, but have a tendency to maximise the difference between the realisations of the individual tones.

Discussion
The results reported in this paper for overall fundamental frequency statistics in the AA and AC speech conditions largely agree with those reported in earlier work. There is a statistically significant raising of mean fundamental frequency in CDS, though the extent of this for the present Cantonese speakers is not as great as has been reported in other languages. Similarly, the range of fundamental frequency variation is increased in CDS for the present speakers, but the increase is by no means as great as has been reported for English. One possible explanation is that, Cantonese being a lexical tone language, there is less freedom for the speaker to manipulate fundamental frequency than there is in intonation languages such as English. The functional distinctiveness of fundamental frequency patterning operates over a much shorter time-scale in a lexical tone language and the points in an utterance where the speaker is constrained by semantic considerations to perform particular movements of fundamental frequency occur much more frequently.

Nevertheless, the overall raising of fundamental frequency and the expansion of the range of variation does take place in Cantonese. There have been explanations in the literature of these CDS tendencies as an attempt on the part of the adult speaker to minimise the child's normalisation task. By adjusting mean fundamental frequency upwards, the difference between what the child is capable of producing and what the child hears is diminished. Other explanations concern the auditory sensitivity of young children. Under this hypothesis the raising of mean fundamental frequency is an attempt to match the frequency at which children show the best response. However, there is evidence (see for instance Kearsley, 1973) that young children show stronger orientating responses to tones at 500 Hz than to tones at 250 Hz. Very few adult speakers are capable of maintaining a mean fundamental frequency of 500 Hz. This hypothesis would also carry the implication that children of the relevant age find it more difficult to understand men than to understand women. There is no independent evidence to support such a conclusion. There is of course the possibility that children's sensitivity to speech-like sounds differs from that in evidence when they are presented with tone stimuli.

Yet another hypothesis put forward to explain CDS fundamental frequency features is that it is an attention attracting device. Werker and McLeod (1989) showed that children aged 0;4-0;5 and 0;7-0;9 spent longer watching a television monitor that presented actors using CDS than when the same actors used adult-directed speech.

Masataka (1992) found that the mean and range of fundamental frequency of Japanese mothers' CDS tended to diverge increasingly from those for adult-directed speech as a function of the number of utterances produced before a response form the infant.

The analysis carried out on the rate of speech of the present subjects in the two speech conditions clearly shows that they employ much slower speech when interacting with the child. This is in accord with results presented in the literature for other languages. There are a number of benefits a young child might obtain from a slower speech rate. One is a gain in processing time. Another is that slower speech is less likely to contain as many occurrences of elision, assimilation and other contextual modifications which are rife in rapid conversational speech between adults. Thus, the child is likely to hear speech which approximates more to an idealized input which avoids variability in the realisation of individual linguistic forms.

The tonal system of Cantonese is a complex one, yet there is evidence that tonal distinctions in this language, as in other lexical tone languages, are acquired before segmental contrasts. Li and Thompson (1977) report this for Modern Standard Chinese (Mandarin). Tse (1977) concludes the same for Cantonese on the basis of a longitudinal study of one child. Another of Tse's findings that is relevant here is that the time span for the acquisition of the production of all 6 Cantonese tones covered a period of only 8 months from 1;2 to 1;9. The majority of the children in the present study fell within this age range. Tse also reports that the perceptual discrimination of lexical tone contrasts began as early as 0;10.

The main finding of the present study regarding Cantonese CDS tone realisation is that there is evidence that adults attempt to maximise the difference between some members of the tonal system. Of course, the overall expansion of the fundamental frequency range employed in CDS must have this effect in absolute terms. However, the fact that differences in tone realisation remain when the fundamental frequency values of the start and end points of the tones are normalised to the speaker's mean frequency. This suggests that the speaker does not rely on a global expansion of the range alone, but locates some members of the system differently within the range in the two speech conditions. This is especially noticeable with the three tones traditionally described as level, but there is also evidence of attempts to increase the distinguishing features of the two rising tones

References
Abberton, E. R. M. & Fourcin A. J. (1984) Electrolaryngography. In C. Code & M. Ball, Eds, Experimental Clinical Phonetics, pp 62-78. London: Croom Helm.

Abberton, E. R. M., Fourcin A. J. & Howard D. M. (1989) Laryngographic assessment of normal voice: A Tutorial. Clinical Linguistics and Phonetics 3, 281-296.

Broen, P. A. (1972) The verbal environment of the language learning child. Monograph of the American Speech and Hearing Association 17

Cheung, K. (1986) The phonology of present day Cantonese. Doctoral dissertation. University of London.

Ferguson, C. A. (1964) Baby talk in six languages. American Anthropologist 66, 103-114

Fernald, A. & Simon, T. (1984) Expanded intonation contours in mothers' speech to newborns. Developmental Psychology 201, 104-113.

Fernald, A., Taeschner, T., Dunn, J., Papousek, M., Boysson-Bardies, B., & Fukui, I (1989) A cross-language study of prosodic modifications in mothers' and fathers' speech to preverbal infants. Journal of Child Language 16, 477-501

Grewel, F. (1959) How do children acquire the use of language? Phonetica 3, 193-202

Kearsley, R. (1973) The newborn's response to auditory stimulation: a demonstration of orienting and defensive behaviour. Child Development 44, 582-590

Kelkar, A. (1964) Marathi baby talk. Word 20, 40-54

Li, C. N. & Thompson S. A. (1977) The acquisition of tone in Mandarin-speaking children. Journal of Child Language 4, 185-200.

Masataka, N. (1992) Pitch characteristics of Japanese maternal speech to infants. Journal of Child Language 19, 213-223.

Norman, J. (1988) Chinese. Cambridge: Cambridge University Press.

Ogle, S. & Maidment, J. A. (1993) Laryngographic analysis of child-directed speech. European Journal of Disorders of Communication 28, 289-297.

Philips, J. (1970) Formal characteristics of speech which mothers address to their young children. Doctoral dissertation, John Hopkins University.

Pye, N. B. & Ratner, C. (1984) Higher pitch in B. T. is not universal: acoustic evidence from Quiche Mayan. Journal of Child Language 11, 515-522

Ramsey, S. R. (1987) The Languages of China. Princeton: Princeton University Press.

Ru±e-DraviÃa, V. (1977) Modifications of speech addressed to young children in Latvian. In C. E. Snow and C. A. Ferguson, Eds, Talking to Children: Language Input and Acquisition, 237-253. Cambridge: Cambridge University Press.

Sachs, J., Brown, R. & Salerno, R. A. (1976) Adults' speech to children. In W. von Raffler-Engel and Y. Lebrun, Eds, Baby Talk and Infant Speech, 240-245. Lisse, Netherlands: Swets & Zeitlinger.

Shute, B. & Wheldall, K. (1989) Pitch alterations in British motherese: some preliminary acoustic data. Journal of Child Language 16, 503-512.

Tse, J. K. P. (1977) Tone acquisition in Cantonese: a longitudinal case study. Journal of Child Language 5, 191-204.

Werker, J. F. & McLeod, P. J. (1989) Infant preference for both male and female infant-directed talk: a developmental study of attentional and affective responsiveness. Canadian Journal of Psychology 43, 230-246.

© 1996 Joanne Siu-Yiu Tang and John A. Maidment

Back to Publications

Back to Phonetics and Linguistics Home Page