AGNĖ PUKEVIČIŪTĖ

Vytautas Magnus University

Research areas: characteristics of spoken Lithuanian.

ASTA KAZLAUSKIENĖ

Vytautas Magnus University

Research areas: studies of spoken Lithuanian: acoustic properties of sounds, phonotactics, stress patterns, norms and usage, speech rhythm; applied linguistics.

CLUSTER ANALYSIS OF TYPOLOGICAL SPEECH RHYTHM GROUPS¹

Cluster Analysis of Typological Groups of Language Rhythm

ABSTRACT

This article aims to answer two questions: 1) Is cluster analysis useful for identifying typological similarities in rhythm, and can it be used to classify languages that have not been analyzed to date? and 2) Which cluster do Lithuanian and Latvian belong to—the syllable-timed or stress-timed language cluster? For the study, five native speakers (a Latvian, a Lithuanian woman, a Swede, an Icelandic woman, and a Chinese woman) translated and recorded the same text, approximately one minute long. The recordings were annotated using the PRAAT program and Paolo Mairano’s

¹ The study was funded by the Research Council of Lithuania (Agreement No. LIT-5-4). Its preliminary results were summarized in Agnė Pukevičiūtė’s master’s thesis (Acoustic Correlates of Rhythm in Lithuanian and Latvian, 2013, supervisor Prof. Asta Kazlauskienė). The authors sincerely thank the informants who recorded and helped segment the research material: Prof. Habil. Dr. Pėteris Vanags, Valgerður Bjarnadóttir, Magnus Liw, Huashan Li, and Associate Professor Dr. Aaušra Saudargienė for assistance in preparing the material for statistical analysis and for advice on working with SPSS.

KEYWORDS: rhythm, stress-timed languages, syllable-timed languages, acoustic correlates of speech rhythm, cluster analysis, Lithuanian, Latvian, Swedish, Icelandic, Chinese.

KEYWORDS: Rhythm, Stress-timed languages, Syllable-timed languages, Acoustic correlates of rhythm, Cluster analysis, Lithuanian, Latvian, Swedish, Icelandic, Chinese.

168

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

Acoustic rhythm correlates were calculated using the program Correlatore (version 2.1). The results of rhythm studies of other languages obtained by the program’s author, Mairano (2011), were also used for the cluster analysis. The data were entered into SPSS, and cluster analysis was performed.

The results of the study show that hierarchical and two-step cluster analyses do not provide unambiguous answers. Hierarchical analysis is fairly well suited to clustering variables that have undergone less statistical processing (ΔV, ΔC,%V, and, in part, varcoΔV and varcoΔC). Two-step analysis should be applied very cautiously when classifying languages according to rhythm, since its results are contradictory.

Lithuanian and Latvian fall into the same cluster for all the variables analyzed, together with many stress-timed languages. Therefore, one could cautiously conclude that the rhythm of both living Baltic languages is similar and determined by stress. They are stress-timed languages.

ANNOTATION

This article aims to answer two questions: 1) whether the cluster analysis is handy for finding typological rhythm similarities and on the basis of it so far unanalysed languages can be classified 2) which cluster – stress-timed or syllable-timed languages – Lithuanian and Latvian languages fall into.

For this study five native speakers (Latvian, Lithuanian, Swedish, Icelandic and Chinese) translated and recorded the same text. Text duration was about 1 minute. Records were annotated with the program PRAAT and acoustic correlates of rhythm were calculated using program Correlatore developed by Paolo Mairano (2.1 version). Results of other languages estimated by the author of the program Mairano (2011) was used for cluster analysis. The data was entered to the SPSS program and a cluster analysis was performed.

The results show that the hierarchical and two-step cluster analysis does not provide unambiguous results. Hierarchical analysis is pretty handy to cluster less statistically processed variables (ΔV, ΔC,%V, partly varcoΔV, varcoΔC). Two-step analysis should be used with caution because of its controversial results.

Lithuanian and Latvian languages from all the variables fall into the same cluster together with a number of stress-timed languages. Therefore, it can be carefully concluded that rhythm of the two Baltic languages is similar and caused by stress. They are stress-timed languages.

1. BRIEF OVERVIEW OF LANGUAGE RHYTHM RESEARCH

Kenneth Lee Pike (1945) vienas iš pirmųjų plačiau aptarė neeiliuotos kalbos ritmiškumo tipologinius skirtumus. Jis skyrė: 1) kirtinio ritmo (ang. stress-timed) kalbas, kurių ritmo pagrindas yra kirtis (pavyzdžiui, anglų, vokiečių kalbos), 2) skiemeninio ritmo (ang. syllable-timed) kalbas, kurių ritmiškumas paremtas

Straipsniai / Articles

169

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

the use of syllables of similar quantity (for example, in French and Spanish). A third group—the mora-timed group—was added somewhat later to these two typological groups; the rhythm of languages belonging to it is based on the mora (for example, Japanese and Tamil).

Many studies from the late twentieth century were based on Pike’s hypothesis of the typological classification of language rhythm. By measuring and comparing the durations of syllables and interstress intervals, researchers sought to substantiate the existence of two rhythm groups and the differences between them (Abercrombie 1967; Roach 1982). Unfortunately, experimental studies did not confirm the isochrony of speech segments—that is, their recurrence at equal intervals or their approximately equal duration. Researchers therefore sought other indicators of linguistic rhythm. Rebecca McKeever Dauer (1983), reasoning that a language’s phonological system directly determines its rhythm, proposed classifying languages into rhythm groups according to the presence or absence of certain phonological features: vowel reduction and syllable-structure complexity. Building on this idea and applying statistical methods, Franck Ramus, Marina Nespor, and Jacques Mehler (1999) attempted to identify acoustic correlates of language rhythm (variables obtained by calculating the durations of vocalic and consonantal intervals): ΔV—the standard deviation of vocalic-interval duration; ΔC—the standard deviation of consonantal-interval duration; and%V—the proportion of vocalic-interval duration. The authors hypothesized that stress-timed languages are characterized by greater variability in vowel duration and more complex syllable structure; therefore, ΔV and ΔC should be higher, and%V lower, than in syllable-timed languages. Unfortunately, the results for the specific languages they analyzed only partly confirmed this.

Later studies drew attention to the possible influence of various linguistic and non-linguistic factors on rhythm. To normalize differences in tempo, Esther Grabe and Ling Ee Low (2002) calculated pairwise variability indices: nPVI for vocalic intervals and rPVI for consonantal intervals. They expected stress-timed languages to be characterized by high variability in vowel and consonant duration, and thus to have high nPVI and rPVI values, while these indices would be low in syllable-timed languages.

Išvengti tempo įtakos trukmei siekė ir Volkeris Dellwo su Petra Wagner (2003). Jie balsinių ir priebalsinių intervalų trukmės standartinį nuokrypį padaugino iš 100 ir apskaičiavo variacijos koeficientus: varcoΔV (balsinių intervalų) bei varcoΔC (priebalsinių intervalų). Tačiau nei poriniai variantiškumo indeksai, nei variacijos koeficientai nepadeda vienareikšmiškai nustatyti ritmo tipologinių panašumų ir suskirstyti kalbų į tipologines klases. Vis dėlto aptarta ritmo parametrų nustatymo metodika gana ilgai buvo, o ir dabar dar tebėra populiari, nors ir susilaukė pagrįstos kritikos (pavyzdžiui, Arvaniti 2012)².

² For a fuller discussion of the language-rhythm research methods reviewed here and their possible application to the description of Lithuanian rhythm, see Kazlauskienė 2013.

170

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

2. RESEARCH METHODOLOGY

This study set out to test a new approach: to experimentally examine the similarity of acoustic rhythm correlates and of speech rhythm more generally by applying cluster analysis. In statistics, this method is used to identify similarities among objects or phenomena and to group them into clusters (i.e., groups of similar objects). The aim of cluster analysis is to group the variables under study so that differences within a cluster are as small as possible, while differences between clusters are as large as possible (Vaitkevičius, Saudargienė 2006).

It was expected that cluster analysis would clarify two matters, which are also the aim of this study.

1. Do the acoustic rhythm correlates of Swedish, Icelandic, and Chinese calculated by Paolo Mairano (2011) and for this study fall into the same clusters? Do the stress-timed Swedish and Icelandic correlates identified by Mairano and by us differ from the syllable-timed Chinese correlate? If these questions are answered in the affirmative, then cluster analysis is indeed useful for identifying typological similarities in rhythm, and languages not yet analysed can be classified on this basis.

2. The second question of interest is which cluster—syllable-timed or stress-timed languages—Lithuanian and Latvian belong to.

The acoustic rhythm correlates of Swedish, Icelandic, and Chinese have already been calculated by other authors (for this study, data were drawn from a publicly accessible online source—Mairano 2011). The first two languages are classified as stress-timed, and Chinese as syllable-timed. The rhythm of Lithuanian and Latvian has not yet been examined using this methodology. Earlier studies of rhythm in Lithuanian treated stress as the basis of rhythm, but measured not the acoustic duration of units, but the distribution of stress across different segments. The rhythm of verse was studied in this way by Juozas Girdzijauskas (1966, 1979) and Leonardas Sauka (1974, 1978), while the rhythm of prose and spoken language was studied by Valerija Ramonaitė (1980, 2006), Bonifacas Stundžia (1991), and Asta Kazlauskienė, Inga Fokaitė (2005).

There had already been indications of rhythmicity in spoken Latvian. Zinny Sans Bond, Dace Markus, and Verna Stockmal (2003) compared the prosody of native Latvian speakers and native Russian speakers speaking Latvian, examining contrastive vowel duration, stress, and rhythm. Their article does not provide a detailed analysis and interpretation of Latvian rhythm, although they calculated the measures proposed by Ramus, Nespor, and Mehler (1999) and by Grabe and Low (2002). Bond, Markus, and Stockmal (2003) assumed in their study that Latvian is characterized by syllable-timed rhythm.

This study uses the standard text proposed by the International Phonetic Association (see http://alt-usage-english.org/north_txt.html) (English The North Wind and the Sun, Lithuanian Šiaurės vėjas ir saulė), which is used not only by rhythm researchers but by all phoneticians. If material of a different kind were analysed, it would be methodologically unsound to compare the results of this study with data from other languages obtained using the aforementioned text.

Straipsniai / Articles

171

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

Five native speakers (a Latvian man, a Lithuanian woman, a Swedish man, an Icelandic woman, and a Chinese woman)³ translated the text and read it aloud. The informants’ speech was recorded on a voice recorder in a quiet room. Each recording is slightly less than one minute long.

The research material was processed and the empirical data extracted in several stages:

1. In consultation with an informant for each language, the recordings were annotated using the PRAAT program. Annotation was based on both the oscillogram display and listening. The completed text annotation grid of the program (ang. TextGride) was saved in *txt format.

2. Using Correlatore (version 2.1), a program developed by Mairano (2011) at Arturo Genre’s experimental phonetics laboratory in Turin, Italy, statistical values for the acoustic rhythm correlates of our recordings were calculated (the results are presented in Table 1). For the cluster analysis, as already mentioned, data from Mairano (2011) were used. These are needed to carry out the cluster analysis properly, since its results are considerably more reliable when a larger amount of data is grouped.

TABLE 1. Statistical values of acoustic rhythm correlates⁴

Language | ΔC |%V | ΔV | varcoΔV | varcoΔC | nPVI | rPVI; Lithuanian | 52 | 54 | 54 | 54 | 66 | 57 | 57; Latvian | 55 | 52 | 58 | 55 | 65 | 70 | 65; Swedish | 55 | 44 | 36 | 39 | 48 | 45 | 68; Swedish 2 | 65 | 49 | 51 | 50 | 66 | 60 | 82; Icelandic | 60 | 40 | 33 | 45 | 56 | 45 | 63; Icelandic 2 | 52 | 47 | 41 | 50 | 60 | 51 | 56; Chinese | 44 | 51 | 44 | 44 | 47 | 48 | 54; Chinese 2 | 48 | 62 | 74 | 51 | 62 | 51 | 53; Japanese | 31 | 51 | 39 | 47 | 40 | 43 | 32; Portuguese | 54 | 49 | 55 | 48 | 44 | 57 | 56; French | 40 | 49 | 40 | 45 | 41 | 45 | 44; Finnish | 47 | 48 | 45 | 47 | 46 | 45 | 55; Greek | 37 | 47 | 34 | 50 | 48 | 52 | 42; German | 66 | 46 | 54 | 56 | 60 | 58 | 65; Estonian | 54 | 46 | 39 | 43 | 51 | 43 | 53; Czech | 58 | 45 | 51 | 51 | 49 | 45 | 61; Turkish | 58 | 45 | 38 | 43 | 55 | 46 | 67

³ All the informants are people with higher education in philology who work or continue their studies at universities: the Latvian man and Icelandic woman are lecturers, the Swedish man is a human resources manager, and the Lithuanian and Chinese women are master’s students.

⁴ Here and throughout, recordings from our study are marked with the index 2, e.g., Swedish 2, whereas Mairano’s data are unindexed.

172

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

Language | ΔC |%V | ΔV | varcoΔV | varcoΔC | nPVI | rPVI Arabic | 55 | 45 | 47 | 56 | 55 | 47 | 64 Italian | 54 | 44 | 44 | 53 | 49 | 48 | 62 Dutch | 46 | 43 | 45 | 62 | 50 | 68 | 53 Danish | 61 | 42 | 40 | 46 | 52 | 50 | 66 Spanish | 52 | 42 | 30 | 42 | 53 | 42 | 60 Russian | 55 | 41 | 34 | 43 | 48 | 45 | 63 English | 59 | 40 | 48 | 62 | 51 | 61 | 70 Polish | 66 | 36 | 32 | 41 | 49 | 41 | 72

3. The data were entered into the SPSS program (for more information, see http://www.insol.lt/statistics/, Vaitkevičius, Saudargienė 2006), and cluster analysis was performed.

Hierarchical and two-step cluster analyses were performed on Lithuanian and Latvian using the SPSS program. In hierarchical analysis, the number of clusters is not known in advance. Clusters are formed according to a hierarchical principle—all data are treated as a single cluster composed of smaller groups. They can be identified either by dividing one overall cluster into smaller ones or by combining small clusters into larger ones. Two-step analysis differs from the method described above in that the number of clusters is known in advance (Vaitkevičius, Saudargienė 2006). Two clusters were selected for this study because it examines two language rhythm groups—stress-timed and syllable-timed. The third group, mora-timed, was not included in the analysis because no hypothesis was proposed that Baltic languages might belong to this group. In cluster analysis, objects are reassigned until similarities within a cluster become considerably greater than similarities between clusters (Vaitkevičius, Saudargienė 2006). The results for Swedish, Icelandic, and Chinese were tested using only two-step cluster analysis, because a small number of objects was selected for examination (six languages). In addition, the number of clusters and the rhythm group to which these languages are assigned are known in advance (two rhythm groups and the results of two studies).

3.

TWO-STEP CLUSTER ANALYSIS OF THE ACOUSTIC CORRELATES OF SWEDISH, ICELANDIC, AND CHINESE

Standard deviation of consonant duration (ΔC) and percentage of vocalic interval duration (%V)

To determine whether the%V and ΔC values for Swedish, Icelandic, and Chinese recorded by Mairano (2011) and in this study are similar and fall into the same group, a

Straipsniai / Articles

173

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

two-step cluster analysis was performed. First, the mean values and standard deviations of the quantitative variables for each cluster were calculated (see Table 2).

TABLE 2. Cluster centers for percentage of vocalic interval duration and standard deviation of consonant duration

%V | ΔC Mean | Standard deviation | Mean | Standard deviation Cluster 1 | 45.00 | 3.916 | 58.00 | 5.715 Cluster 2 | 56.50 | 7.778 | 46.00 | 2.828 Cluster overall | 48.83 | 7.521 | 54.00 | 7.720

The table shows that the mean ΔC value for the first cluster is higher than that for the second, whereas the reverse is true for the variable%V—the second cluster has a higher value, although in this case the standard deviation is also quite large. Thus, the first cluster is characterized by a more complex syllable structure (i.e., more consonants per syllable (higher ΔC)), which also results in a lower percentage of vocalic intervals (%V).

The results of this analysis show that the most similar%V and ΔC values, forming one cluster, are those of the stress-timed languages Icelandic (%V = 40, ΔC = 60), Icelandic 2 (%V = 47, ΔC = 52), Swedish (%V = 44, ΔC = 55), and Swedish 2 (%V = 49, ΔC = 65), while the syllable-timed Chinese (%V = 51, ΔC = 44) and Chinese 2 (%V = 62, ΔC = 48) samples formed another cluster. Thus,%V and ΔC are fairly reliable indicators showing that the languages analyzed do indeed differ, and are independent of objective factors relating to researchers or informants (speech rate, intonation, etc.) and subjective factors (recording quality, annotation differences, etc.).

Variation coefficients

There is also a fairly substantial difference between the centers of the two clusters in terms of variation coefficients. Table 3 shows that the means of varcoΔV and varcoΔC are higher for the second cluster.

TABLE 3. Cluster centers by variation coefficients

varcoΔV varcoΔC Mean Standard deviation Mean Standard deviation Cluster 1 42,67 3,215 50,33 4,933 Cluster 2 50,33 0,517 62,67 3,055 Cluster Overall 46,50 4,680 56,50 7,688

174

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

Unfortunately, this time the results are highly unexpected: the most similar values of varcoΔV and varcoΔC are found in the languages analyzed by Mairano (2011) that also form the first cluster: Icelandic (varcoΔV = 45, varcoΔC = 56), Swedish (varcoΔV = 39, varcoΔC = 48), and Chinese (varcoΔV = 44, varcoΔC = 47). All of our recordings fall into the second cluster: Icelandic 2 (varcoΔV = 50, varcoΔC = 60), Swedish 2 (varcoΔV = 50, varcoΔC = 66), and Chinese 2 (varcoΔV = 51, varcoΔC = 62). Thus, the variation coefficient, which was supposed to provide a more objective basis for the typological classification of languages, creates more confusion and suggests that its results depend heavily on the annotation experience specific to each language. The variation coefficients of our recordings are similar, which suggests that when annotating, it may not be sufficient merely to consult native-speaker informants; the annotation itself must be carried out by native-speaker phoneticians.

Pairwise variability indices

The difference between the centers of the two clusters in terms of nPVI and rPVI is fairly substantial (see Table 4).

⁴ TABLE. Cluster centers by pairwise variability indices

nPVI

rPVI

Mean

Standard deviation

Mean

Standard deviation

Cluster

1

48,00

3,000

58,80

6,458

2

60,00

.

82,00

.

Overall

50,00

5,586

62,67

11,094

However, the grouping of languages is even more ambiguous than when examining the coefficients of variation. The first cluster comprises all languages except Swedish 2, analyzed for this study (its rPVI is very high—82, although its nPVI is similar to those of the other languages—60); it was the only language to fall into the second cluster. The pairwise variability indices of the other languages are very similar: Icelandic (nPVI = 45, rPVI = 63), Icelandic 2 (nPVI = 51, rPVI = 56), Chinese (nPVI = 48, rPVI = 54), Chinese 2 (nPVI = 51, rPVI = 53), Swedish (nPVI = 45, rPVI = 68). In this case, it is evident that a two-step cluster analysis is not useful, since the first cluster needs to be divided into smaller groups. To summarize the results of the two-step cluster analysis of the three languages, it can be said that the standard deviation of consonant duration (ΔC) and the percentage of vocalic interval duration (%V) are fairly reliable indicators for grouping languages into classes. In this case, two-step cluster analysis using coefficients of variation and pairwise variability indices raises more questions than it answers.

Straipsniai / Articles

175

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

4. THE RHYTHM OF LITHUANIAN AND LATVIAN BASED ON CLUSTER ANALYSIS

Standard deviations of duration (ΔC, ΔV) and the percentage of vocalic interval duration (%V)

To determine which typological rhythm group Latvian and Lithuanian belong to, hierarchical cluster analysis was first conducted to examine how languages group according to ΔV,%V, and ΔC, and which cluster Lithuanian and Latvian fall into. Data on other languages, presented in Mairano’s work (2011: 170, 179, 199), were also used for this analysis.

The results of the hierarchical cluster analysis show that Lithuanian, Latvian, German, Arabic, English, Czech, Italian, and Portuguese have the most similar values for ΔV,%V, and ΔC, and therefore form a single cluster, in which ΔV ranges from 44 to 58,%V from 40 to 54, and ΔC from 52 to 66. In this cluster, Italian alone is considered closer to syllable-timed rhythm (as can be seen, the cluster analysis seems to cast doubt on this), while all the other languages (except, of course, the Baltic languages under analysis, whose rhythm type is unknown) are stress-timed languages.

The second group was formed by many syllable-timed languages, but it also included a considerable number of languages generally considered stress-timed or mixed-rhythm languages (see Table 5). For these languages, ΔV ranges from 30 to 45,%V from 36 to 51, and ΔC from 37 to 66. One thing is certain from these results: not a single prototypical syllable-timed language (French, Spanish, Greek) fell into the first cluster, while only two prototypical stress-timed languages (Russian, Dutch) fell into the second. As Table 5 shows, Lithuanian and Latvian fall into the first cluster and are more similar to prototypical stress-timed languages (data on the acoustic correlates of the languages are presented in Table 1).

TABLE 5. Results of hierarchical cluster analysis based on standard deviations and vowel proportion⁵

Cluster number

Rhythm groups and languages

Stress-timed rhythm

Syllable-timed rhythm

Undetermined

1

English, German, Arabic, Portuguese, Czech

Italian

Lithuanian, Latvian

2

Russian, Dutch, Icelandic, Turkish, Danish, Polish, Swedish

Spanish, French, Greek, Chinese, Estonian, Finnish

⁵ In this and the other tables, the prototypical languages of a specific rhythm group are highlighted in bold; the rhythmic classification of the other languages is debated (some of them are even considered either unclassifiable or mixed-rhythm).

176

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

The results of the hierarchical analysis of the three variables lead to the conclusion that prototypical stress-timed and syllable-timed languages fall into different clusters. Thus, according to the features analyzed, the languages in these groups differ, but whether these features are indicators of different language rhythms is a question for another study.

After conducting a two-step cluster analysis (the number of clusters selected was, naturally, two), it can be seen that the mean ΔV of the languages in the first cluster is higher, while that of the second cluster is lower (see Table 6). The languages in the first cluster are characterized by greater variability in the duration of vocalic intervals. The mean value of the ΔC variable is lower in the first cluster and somewhat higher in the second. However, the difference is not very large, especially considering that the data in the first cluster are highly dispersed (the standard deviation is high—8.15).

TABLE 6. Cluster centers based on standard deviations and vowel proportion

%V | ΔV | ΔC Mean | Standard deviation | Mean | Standard deviation | Mean | Standard deviation Cluster | 1 | 47.75 | 3.388 | 47.58 | 6.999 | 50.67 | 8.150 Cluster | 2 | 41.78 | 3.032 | 36.67 | 5.408 | 57.78 | 4.295 Overall | 45.19 | 4.377 | 42.90 | 8.324 | 53.71 | 7.544

The results of the two-step cluster analysis, presented in Table 7, show that the ΔV,%V, and ΔC values for Lithuanian and Latvian are most similar to those of the stress-timed languages German, Arabic, Dutch, Portuguese, and Czech, and the syllable-timed languages Italian, Greek, Chinese, Finnish, and French. The first cluster, which according to the data of this study also includes the Baltic languages, consists of 12 languages (57%); the second consists of 9 (43%).

TABLE 7. Results of the two-step cluster analysis based on standard deviations and vowel proportion

Cluster number | Rhythm groups and languages Stress-timed rhythm | Syllable-timed rhythm | Undetermined 1 | Arabic, German, Dutch, Portuguese, Czech | French, Greek, Italian, Chinese, Finnish | Lithuanian, Latvian 2 | English, Russian, Icelandic, Turkish, Danish, Polish, Swedish | Spanish, Estonian

It can be seen that the results of this analysis differ from those of the hierarchical analysis in that more syllable-timed languages fell into the same cluster together with Lithuanian and Latvian

Straipsniai / Articles

177

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

languages: Italian, Chinese, Finnish, and even the prototypical French and Greek languages. Another prototypical syllable-timed language (Spanish) fell into the second cluster. This time, the prototypical stress-timed languages were also split: English into the second cluster and German into the first. Thus, when analyzing three variables (the most important indicators of the typological class of language rhythm), hierarchical rather than two-step cluster analysis is more suitable.

Variation coefficients

Hierarchical cluster analysis using the variables varcoΔV and varcoΔC shows that Lithuanian, Latvian, and four prototypical stress-timed languages—Arabic, Dutch, English, and German—fall into one cluster (see Table 8). The varcoΔV values of these languages range from 54 to 62, and their varcoΔC values from 50 to 66. The languages in the second cluster have varcoΔV values ranging from 39 to 53 and varcoΔC values ranging from 41 to 56.

TABLE 8. Results of hierarchical and two-step cluster analyses based on variation coefficients

Cluster number

Rhythm groups and languages

Stress-timed rhythm

Syllable-timed rhythm

Undetermined

1

English, Dutch, German, Arabic

–

Lithuanian, Latvian

2

Russian, Portuguese, Icelandic, Turkish, Danish, Czech, Polish, Swedish

Spanish, French, Greek, Estonian, Italian, Chinese, Finnish

As can be seen from the results of the two-step cluster analysis, the means of both variation coefficients in the first cluster are higher than those in the second (see Table 9).

TABLE 9. Cluster centers based on variation coefficients

varcoΔV

varcoΔC

Mean

Standard deviation

Mean

Standard deviation

Cluster

1

57,50

3,564

57,83

6,911

2

45,33

3,885

49,07

3,955

Total

48,81

6,743

51,57

6,274

The results of the hierarchical and two-step cluster analyses are identical. The clusters comprise the same languages: one includes Lithuanian, Latvian, German, Arabic, English, and Dutch, while the other comprises the remaining languages (see Table 8).

178

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

Hierarchical and two-step cluster analysis of variability coefficients clearly shows the similarity between the rhythm of Lithuanian, Latvian, and stress-timed languages, and fairly accurately separates prototypical stress-timed languages from syllable-timed ones.

Pairwise variability indices

Hierarchical cluster analysis based on pairwise variability indices yields highly debatable results (see Table 10). All languages except French and Greek fall into a single cluster. The rPVI values of the languages in the first cluster range from 53 to 72, and their nPVI values range from 41 to 70. The rPVI values of the languages in the second cluster range from 42 to 44, and their nPVI values range from 45 to 52.

TABLE 10. Results of hierarchical cluster analysis based on pairwise variability indices

Cluster number | Rhythm groups and languages | Stress-timed rhythm | Syllable-timed rhythm | Undetermined | 1 | English, Dutch, German, Arabic, Russian, Portuguese, Icelandic, Turkish, Danish, Czech, Polish, Swedish | Spanish, Estonian, Italian, Chinese, Finnish | Lithuanian, Latvian | 2 | – | French, Greek

Table 11 presents the results of the two-step cluster analysis. It is noteworthy that the mean rPVI values of the two clusters are almost identical, which casts serious doubt on the reliability of this measure (particularly given the high standard deviation).

TABLE 11. Cluster centers based on pairwise variability indices

nPVI | rPVI | Mean | Standard deviation | Mean | Standard deviation | Cluster | 1 | 61,83 | 5,776 | 61,00 | 6,603 | 2 | 45,80 | 2,883 | 59,60 | 8,534 | Total | 50,38 | 8,322 | 60,00 | 7,893

The first cluster consists of six languages (29%), and the second of 15 (71%) (see Table 12). Compared with the results of the hierarchical cluster analysis, we find that German, English, Dutch, and Portuguese remain in the first cluster alongside Lithuanian and Latvian. All syllable-timed languages fall into the second cluster.

Straipsniai / Articles

179

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

TABLE 12. Results of two-step cluster analysis based on pairwise variability indices

Cluster number

Rhythm groups and languages

Stress-timed rhythm

Syllable-timed rhythm

Undetermined

1

English, Dutch, German, Portuguese

–

Lithuanian, Latvian

2

Russian, Arabic, Icelandic, Turkish, Danish, Czech, Polish, Swedish

Spanish, French, Greek, Estonian, Italian, Chinese, Finnish

The results of the two-step cluster analysis using nPVI and rPVI as variables indicate a rhythmic similarity between Lithuanian, Latvian, and stress-timed languages and, like the analysis of variability coefficients, clearly distinguish prototypical stress-timed languages from syllable-timed ones.

5. CONCLUSIONS

To summarize the attempt to apply cluster analysis methods to identify typological rhythm classes, the following conclusions can be drawn.

Cluster analysis (both hierarchical and two-step) does not yield unequivocal results. Hierarchical analysis is quite suitable for clustering variables that have undergone less statistical processing (ΔV, ΔC,%V, and partly varcoΔV and varcoΔC); two-step analysis may perhaps also be used with more complex indices (PVI). Nevertheless, in this case, two-step analysis raises more doubts, if only because just two possibilities have been selected (either stress-timed or syllable-timed rhythm; it cannot be mixed).

Lithuanian and Latvian fall into the same cluster for all variables, together with many stress-timed languages. Therefore, it would be possible to draw the cautious conclusion that the rhythms of both Baltic languages are similar and are determined by stress. These are stress-timed languages.

Cluster analysis (especially the comparison of data from two different studies of Swedish, Icelandic, and Chinese) showed that annotation precision is particularly important for the study of rhythm. On the other hand, this is only a pilot study, so these conclusions are relevant not because of the typological classification of language rhythm itself or measures of language rhythmicity. They may serve as guidelines for further research or provide another reason to question the existence of strictly defined language rhythm classes.

LITERATŪRA

A b e r c r o m b i e D a v i d 1967: Elements of general phonetics. Edinburgh: Edinburgh University Press.

A r v a n i t i A m a l i a 2012: The usefulness of metrics in the quantification of speech rhythm. Journal of Phonetics 40, 351–373.

180

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

Bond Zinny Sans, Markus Dace, Stockmal Verna 2003: Prosodic and Rhythmic Patterns Produced By Native and Non-native Speakers of a Quantity-Sensitive Language. 15th International Congress of Phonetic Sciences. Barcelona, 527–530.

Dauer Rebecca McKeever 1983: Stress-timing and syllable-timing reanalysed. Journal of Phonetics 11, 51–62.

Dellwo Volker, Wagner Petra 2003: Relationships between language rhythm and speech rate. Proc. of the International Congress of Phonetic Science, 471–474.

Girdzijauskas Juozas 1966: Lietuvių eilėdara. Silabinės-toninės sistemos susiformavimas. Vilnius: Vaga.

Girdzijauskas Juozas 1979: Lietuvių eilėdara: XX a. Vilnius: Mokslas.

Grabe Esther, Low Ee Ling 2002: Durational Variability in Speech and the Rhythm Class Hypothesis. Papers in Laboratory Phonology 7, 1–16.

Kazlauskienė Asta, Fokaite Inga 2005: Dėl lietuvių kalbos teksto prozodijos. Žmogus ir žodis 7(1), 35–38.

Kazlauskienė Asta 2013: Kai kurie akustiniai lietuvių kalbos ritmiškumo parametrai. Baltistica 48(1), 19–36.

Mairano Paolo 2011: Rhythm Typology: Acoustic and Perceptive Studies. Doctoral dissertation. Turin: University of Turin.

Pike Kenneth Lee 1945: The Intonation of American English. Ann-Arbor: University of Michigan Press.

Pukevičiūtė Agnė 2013: Akustiniai lietuvių ir latvių kalbų ritmo koreliatai. Magistro darbas. Kaunas: Vytauto Didžiojo universitetas.

Ramonaite Valerija 1980: Prozos kalbos ritmas. Konstruktyvieji faktoriai ir pagalbiniai formantai. Literatūra ir kalba 16, 38–65.

Ramonaite Valerija 2006: Pirminis Juozo Baltušio „Parduotų vasarų“ ritmas. Acta humanitarica universitatis Saulensis, 307–333.

Ramus Franck, Nespor Marina, Mehler Jacques 1999: Correlates of linguistic rhythm in the speech signal. Cognition 73(3), 265–292.

Roach Peter 1982: On the Distinction between ‘Stress-timed’ and ‘Syllable-timed’ Languages. – David Crystal. Linguistic controversies. London: Edward Arnold, 73–79.

Sauka Leonardas 1974: Lietuvių folklore eilet yros ir teorijos klausimai. – Literatūra ir kalba 13, 192–207.

Sauka Leonardas 1978: Lietuvių liaudies dainų eilėdara. Vilnius: Vaga.

Stundžia Bonifacas 1991: Kirtis tekste. Žodžiai ir prasmės 1, 86–92.

Vaitkevičius Raimundas, Saudargienė Aušra 2006: Statistika su SPSS psichologiniuose tyrimuose: mokomoji knyga. Kaunas: VDU leidykla.

Vaitkevičius Raimundas, Saudargienė Aušra 2012: Daugiamačiai duomenų analizės metodai psichologijoje. Kaunas.

Straipsniai / Articles

181

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

PRIEDAS

1. Tiriamieji tekstai lietuvių kalba

Vieną dieną Šiauriniam vėjui ir Saulei besiginčijant, kuris iš jų stipresnis, pasirodė keliautojas, vilkintis apsiaustą. Tada jie susitarė, kad tas, kuris privers keliautoją nusivilkti apsiaustą, bus laikomas stipresniu.

Pirmiausia Šiaurinis vėjas pūtė taip stipriai, kiek tik galėjo, bet kuo stipriau jis pūtė, tuo tvirčiau keliautojas apsigaubdavo apsiaustą; galiausiai Šiaurinis Vėjas pasidavė.

Tada Saulė pradėjo kaitinti ir keliautojas tuoj pat nusivilko apsiaustą. Taip Šiaurinis vėjas turėjo pripažinti, kad Saulė stipresnė už jį.

2. Tiriamieji tekstai latvių kalba

Vienu dienu Ziemeļu vējam un Saulei strīdoties par to, kurš no viņiem ir stiprāks, nāca ceļotājs, ietīnies mētelī. Viņi savā starpā vienojās, ka tas, kurš pirmais liks ceļotājam novilkt savu mēteli, tiks uzskatīts par stiprāko.

Ziemeļvējš pirmais pūta tik stipri, cik vien varēja, bet, jo stiprāk viņš pūta, jo stingrāk ceļotājs ietinas savā mētelī; galu galā Ziemeļvējš tomēr padevās.

Tad Saule sāka stipri sildīt un ceļotājs uzreiz novilka savu mēteli. Tā Ziemeļvējam bija jāatzīst, ka Saule bija stiprāka par viņu.

3. Tiriamieji tekstai švedų kalba

Nordanvinden och Solen tvistade en gång om vem av dem som var starkast. Just då kom en vandrare vägen fram insvept i en varm kappa. De kom då överens om att den som först kunde få vandraren att ta av sig kappan, han skulle anses vara starkare än den andra.

Då blåste nordanvinden så hårt han nånsin kunde, men ju hårdare han blåste desto tätare svepte vandraren kappan om sig, och till slut gav Nordanvinden upp försöket.

Då lät Solen sina strålar skina helt varmt och genast tog vandraren av sig kappan och så var Nordanvinden tvungen att erkänna att solen var den starkaste av de två.

4. Tiriamieji tekstai islandų kalba

Dag einn þegar Norðvindurinn og Sólin voru að rífast um það hvor þeirra væri sterkari, birtist ferðamaður vafinn inn í frakka. Þeir voru sammála um að sá sem gæti blásið frakkanum af ferðamanninum myndi teljast sterkari en hinn.

182

Acta Linguistica Lithuanica LXIX

Klasterinė tipologinių kalbos ritmo grupių analizė

Norðanvindurinn blés eins fast og hann gat, en því fastara sem hann blés, því þéttara varð ferðamaðurinn kápunni um sig, og að lokum gafst Norðanvindurinn upp á að reyna.

Þá fór Sólin að skína, og undireins fór ferðamaðurinn úr frakkanum. Þá varð Norðanvindurinn að viðurkenna að Sólin væri sterkari en hann var.

6. Tiriamieji tekstai kinų kalba

You yi tian, dang yi ge guo zhe da yi de lv xing zhe chu xian de shi hou, bei feng he tai yang zheng zai zheng chao jiu jing shui geng jiang. Ta men tong yi shui xian rang lv xing zhe ba wai tao tuo diao jiu ren wei shui bi ling yi ge qiang.

Sui hou, bei feng jiu jing ta zui da de nu li lai chui feng, dan shi, ta chui de yue da, lv xing zhe jiu ba wai tao guo de yue jin. Zui hou,bei feng fang qi le.

Jie xia lai, tai yang kai shi fa chu re liang, lv xing zhe li ji jiu tuo xia le wai tao. Suo yi, bei feng bu de bu cheng ren tai yang bi ta geng qiang.

Cluster Analysis of Typological Groups of Language Rhythm

SUMMARY

This article aims to answer two questions: 1) whether the cluster analysis handy finding rhythm typological similarities, and it can be classified on the basis of it so far unanalysed languages 2) in which a cluster – stress-timed or syllable-timed languages – Lithuanian and Latvian languages fall into.

Standard text, proposed by the International Phonetics Association (see http://alt-usage-english.org/north_txt.html) (Eng. The North Wind and the Sun, Lith. Šiaurės vėjas ir saulė) was selected for this study. Five native speakers (Latvian, Lithuanian, Swedish, Icelandic and Chinese) translated and recorded it (each record was much shorter than 1 min.). The records were annotated by program PRAAT and acoustic correlates of rhythm of records that we have done have been calculated using program Correlator (2. 1 version) which was developed by Pado Mairano (2011) in Arturo Genre experimental phonetics laboratory of Turin (Italy). It was found that Lithuanian ΔC=52,%V=54, ΔV=54, varcoΔV=54, varcoΔC=66, nPVI=57, rPVI=57; Latvian ΔC=55,%V=52, ΔV=58, varcoΔV=55, varcoΔC=65, nPVI=70, rPVI=65; Swedish ΔC=65,%V=49, ΔV=51, varcoΔV=50, varcoΔC=66, nPVI=60, rPVI=82; Icelandic ΔC=52,%V=47, ΔV=41, varcoΔV=50, varcoΔC=60, nPVI=51, rPVI=56; Chinese ΔC=48,%V=62, ΔV=74, varcoΔV=51, varcoΔC=62, nPVI=51, rPVI=53.

Straipsniai / Articles

183

AGNĖ PUKEVIČIŪTĖ, ASTA KAZLAUSKIENĖ

Data of other languages estimated by the author of the program Mairano (2011) was used for cluster analysis. The data was entered to the SPSS program and a cluster analysis was performed.

The results show that the cluster analysis (both hierarchical and two-steps) does not provide unambiguous results. Hierarchical analysis is pretty handy to cluster less statistically processed variables (ΔV, ΔC,%V, partly varcoΔV, varcoΔC), a two-step analysis might be possible to rely on, and with more complex indices (PVI). However, a two-step analysis is begging the question, because it does not leave the third option (stress-timed or syllable-timed, cannot be mixed).

Lithuanian and Latvian languages according to all the variables fall into the same cluster together with a number of stress-timed languages. Therefore, it can be carefully concluded that rhythm of the two Baltic languages is similar and caused by stress. They are stress-timed languages.

Cluster analysis (in particular, Swedish, Icelandic, Chinese Language comparison between the two types of data) has shown that the annotation precision is particularly important in the researches of rhythm. On the other hand, the findings are relevant not for the typological classification of languages’ rhythm and not for the indicators of languages’ rhythm. They can be as guidelines of further research or another excuse to question existence of strictly defined classes of speech rhythm.

Įteikta 2013 m. liepos 11 d.

AGNĖ PUKEVIČIŪTĖ

Vytauto Didžiojo universitetas

Kristijono Donelaičio g. 58, LT–44248 Kaunas, Lietuva

[email protected]

ASTA KAZLAUSKIENĖ

Vytauto Didžiojo universitetas

Kristijono Donelaičio g. 58, LT–44248 Kaunas, Lietuva

[email protected]

184

Acta Linguistica Lithuanica LXIX