GRToBI 

(GREEK TONES AND BREAK INDICES)

Home    Publications    CV    Contact    Speech Lab    Linguistics    UCSD      

developed by

Amalia Arvaniti
&
Mary Baltazani
 

Please note:
(a) this page is perpetually under construction, so we would appreciate it if you would notify us of any mistakes, inconsistencies or problems;
(b) the most complete description of GRToBI currently available is Arvaniti & Baltazani, 2005
(c) this page provides downloadable files, and includes some updates to the system which it was not possible to include in Arvaniti & Baltazani (2005),

but it is by necessity brief and in no way could be seen as a substitute of the published papers on GRToBI.

What is GRToBI?

GRToBI is a tool for the annotation of Greek spoken corpora; it provides a system for annotating intonational, prosodic and (limited) phonetic information,

though users can add tiers that encode other types of information as well. Although it was originally designed to work on Waves+, the audio and annotation

files have been converted so that they can now be read in PRAAT.  You can download them from this site.

GRToBI is not a transcription system for Greek intonation, i.e. it is not equivalent to a list of IPA symbols for Greek intonation. This is so for two related reasons.

First, GRToBI assumes a particular view of prosody and of the relationship between phonetics and phonology. In particular, GRToBI is based on the

autosegmental-metrical framework of intonational phonology which assumes a principled distinction between phonetics and phonology. The analysis on which GRToBI

is based is phonological; that is, GRToBI is not meant to be a surface phonetic transcription of Greek intonation in other words. Thus, the annotation labels may not always

be phonetically transparent, and are not intended to capture all possible variations in phonetic realization. On the other hand, the differences indicated by the autosegmental

representations are meant to be meaningful, that is to reflect pragmatic differences related to the melody. Second,  GRToBI has been designed to represent

the prosodic system of Greek as spoken in Athens. Thus the tonal inventory may include phonological entities (e.g. pitch accents) not attested in other varieties

of Greek and may not include entities that are attested in those varieties; further, the phonetic realization of the same entity may differ between the Athenian

variety assumed here and other varieties; finally, it is quite possible that certain configurations (e.g. the combinations of  phrasal tones described below) do not

occur in all varieties of Greek or if they do occur they have different pragmatics.

 

In terms of design, GRToBI is similar to the original ToBI system for American English or MAE ToBI (MAE stands for Mainstream American English; see Silverman

et al., 1992; Beckman et al. 2005). GRToBI has been adapted from this original design so that prosodic phenomena requiring special attention in Greek, such as sandhi,

can be more easily annotated. 

 

The prosodic and intonational analyses assumed in GRToBI have been based
(a) on  existing research on various aspects of Greek prosody (see Bibliography); since new findings are continuously added to this body of research, it is natural

that particular aspects of the phonological analysis assumed in GRToBI may be revised in the light of new evidence.
(b) on the transcription of a corpus of spoken Standard Greek that includes data from several speakers using a variety of styles (read text, news broadcasting,

interviews, spontaneous speech).

Uses of GRToBI

GRToBI is being used to develop a publicly available corpus of annotated utterances. Prosodically annotated corpora are an important language resource for

several reasons. Corpora based research can contribute to the better understanding of prosody (see e.g. Arvaniti & Pelekanou 2002), a neglected aspect of

spoken language, the importance of which in speech production, speech perception and language acquisition has now begun to emerge. A better understanding

of prosody will also lead to more natural speech synthesis and, possibly, more efficient speech recognition systems.

 

Further, GRToBI is based on an analysis of Greek prosodic structure. This analysis, developed on the basis of existing research and the annotation of the GRToBI

corpus itself, is the first systematic description of Greek prosody and, as such, useful for theoretical reasons.

Thus, the GRToBI  database and prosodic analysis system can be used for several purposes:

by searching the corpus;

GRToBI annotated utterances as examples;

The GRToBI annotation system

A GRToBI annotated file consists of an audio recording of the utterance (in wav format) and an annotation file (currently in the PRAAT textgrid format).

When the wav and textgrid files are opened together the following appear: the waveform of the utterance, the spectrogram and pitch contour of the utterance

(in Hz), as well as the five GRToBI annotation tiers (see Figure 1 for an example):

The Tone Tier

For the intonational analysis of Greek utterances we recognize three types of tonal events, pitch accents, phrase accents, and boundary tones, and two levels

of phrasing, the intermediate phrase (ip) and the intonational phrase (IP).
 

The pitch accents

A pitch accent is a melody that is phonologically associated with a metrically strong syllable; phonetically, a pitch accent co-occurs (more or less) with the stressed

syllable it is phonologically associated with.  Pitch accents are “prominence cueing” (a term coined by Francis Nolan) in that they indicate that the syllable with

which they co-occur is meant to be construed as being metrically prominent (i.e. stressed). In Greek there is typically one prominent syllable in each content word;

its position is lexically specified. Most function words do not have prominent syllables (though they may carry orthographic accent), but some do, such as the word

[katá] when it means ‘against’. In some cases, Greek words may carry two pitch accents: this happens when a word is stressed on the antepenult and is followed

by an enclitic (e.g. [to aftokínitó mu] ‘the car my’), or when a word is stressed on the penult and is followed by two enclitics (e.g. [fére mú to] ‘bring me it’). In such

cases, both stressed syllables can be accented; if there is only one accent, this falls on the rightmost stressed syllable of the entire group. Thus, in Greek we can

distinguish  between unstressed syllables, such as [re] in [fére mú to], stressed but unaccented syllables, such as [fe] in [fere mú to], and stressed and accented

syllables, such as [fé] and [mú] in [fére mú to].

Our research suggests that in Greek we can distinguish five pitch accents: H*, L*, L*+H, L+H*, H* and H*+L. The typical distribution and phonetic realization of

these accents is described below and shown in Figures 1, 2, 3, and 4. Each utterance has at least one pitch accent, but typically more, since in Greek most stressed

syllables are also accented; thus, while an English phrase such as Mary loves John is most likely uttered without an accent on loves, in a similar Greek phrase, such

as [i ma'ria mi'lai sto 'mano] ‘Maria is talking to Manos’, all three content words are accented. The last accent of an utterance is called the nucleus. By default,

the nucleus falls on the last content word of an utterance (if it is a declarative—for other types of utterance, see below), but it may occur on an earlier word under

narrow or contrastive focus.

whereas H* signals broad focus (Baltazani 2003). The H* accent lacks the initial dip associated with the L tone of the L+H* (Arvaniti et al. 2006b) and its peak is probably aligned

earlier in the accented vowel, though quantitative data on this point are not yet available.

Baltazani & Jun 1999; Baltazani 2007b), and in the “suspicious” calling contour (that is, when the vocative is to be interpreted as “is that you?”)

L tone is aligned at or slightly before the onset of the accented syllable, and the H tone is aligned at the beginning of the first post-accentual vowel (Arvaniti & Ladd, 1995;

Arvaniti et al. 1998). The realization of L*+H is different in contexts showing tonal crowding (Arvaniti et al. 2000).

lies  in the alignment of the H tone: the H tone of L+H* is well within the accented vowel, whereas the H tone of the L*+H aligns early in the first post-accentual vowel.

throughout this syllable (see the last accent in Figure 9). In terms of meaning, the use of H*+L conveys a sense of “stating the obvious” that is, the implication that the addressee

should have known or expected the answer.

At this stage, it is not clear whether there is any particular meaning associated with downstep, or with particular downstepped accents.

The phrase accents

In GRToBI, three phrase accents, H-, L- and !H-, are assumed to exist in Greek. As mentioned earlier, our analysis suggests that Greek has two levels of phrasing,

the intermediate and intonational phrase (ip and IP respectively), and it is assumed here that phrase accents demarcate the right boundary of intermediate phrases.

This analysis is based on the following observations. Tones associated with ips typically show simple F0 movements, unlike those associated with IPs which can show complex

pitch configurations. Further there is a difference in scaling between ips and IPs, in cases where both have similar pitch movements, with ip boundaries exhibiting less extreme

F0 values than IP boundaries (i.e. a H- is scaled lower than a H-H% configuration). In addition, the pauses after IPs (even non-final ones) are longer and more frequent

than those for ips. Recent research also suggests that left ip and IP boundaries are associated with “prosodic strengthening” manifested as lengthening of ip and IP initial

consonants (Kastrinaki 2003). Non-final intermediate phrases typically have a H- or L- phrase accent at their right edge.  !H-, on the other hand, is used only in certain types

of stylized intonation and then only in utterance-final ips (i.e. it is always followed by a boundary tone).

More recently, Arvaniti et al. (2006a) have shown that the melody of Greek polar questions is difficult to accommodate with this inventory of phrase accents, and suggest

that the phrase accent of these questions is a bitonal L+H-. In addition, this phrase accent does not always co-occur with the right edge of the intermediate phrase it is

associated with; rather, when the nucleus of the polar question is on a non-final word, the L+H- phrase accent aligns with the last stressed syllable of the question. The

two patterns of alignment of the L+H- used in Greek polar questions are illustrated in Figure 4.

The boundary tones

GRToBI includes three types of boundary tone, H%, L%, and !H%. These boundary tones demarcate the right edges of intonational phrases. They combine with most phrase

accents into configurations which are frequently interpreted in the ways shown below (though the list is only indicative; the interpretation of a contour  depends also on

the utterance and the context in which it is used). 

Possible combinations of phrase accent and boundary tone and their usages

L-L%

declaratives, negative declaratives, imperatives, wh-questions 

L-H%

“involved” continuation rise, “suspicious” calls

H-L%

yes-no questions, requesting calling contour (note: according to Arvaniti et al. 2006a this combination is L+H-L%)

H-H% 

continuation rise, questioning calling contour

L-!H% 

“involved” wh-questions, negative declaratives (showing reservation),  requesting imperatives

H-!H% 

stylized continuation rise

!H-!H%

stylized call, incredulous questions

!H-H%

polite stylized call

The Prosodic Words Tier

The Prosodic Words (PrWords) Tier is a phonetic transcription using ASCII characters (see Appendix Ifor coventions). This tier facilitates the analysis of sandhi

(connected speech phenomena, such as segment assimilations and deletions across word boundaries), and fast speech rules, by encoding their outcome. Like all

transcriptions, this tier has its limitations, and is not meant to be a substitute for acoustic analysis; rather, it allows annotators to flag instances of sandhi for

more detailed acoustic analysis. The PrWords Tier provides information about stress, since this information cannot be deduced from the transliteration in the

Words Tier or derived from a dictionary (e.g., as mentioned, content Greek monosyllables, as well as some function words, are normally stressed and pitch accented

in speech, but not in orthography; in contrast, disyllabic function words are orthographically accented, but most do not normally carry stress in speech). In this tier,

each prosodic word (defined as a sequence of items showing total cohesion) is transcribed as one label.
 

The Words Tier

This tier is a transliterated version of the text, equivalent to the Orthographic Tier in the American English ToBI  (see Appendix II for conventions).
 

The Break Index Tier (and its diacritics)

There are four break indices, 0, 1, 2, and 3. Break indices mark cohesion (or the lack thereof) between constituents in an utterance.

that may bear only one pitch accent (with the noted exception of hosts and clitics with two accents). Several types of sandhi may occur across a BI 0 boundary,

however, sandhi is not necessary for a BI 0 to be used. For example, a proclitic particle like /na/ and the following verb are perceived as one PrWord by native

speakers, but no sandhi can occur between /na/ and a consonant-initial verb.

all PrWords following an early focus are de-accented; Baltazani & Jun, 1999; Botinis, 1998). In general, if an item is accented, then it should be considered

a separate PrWord. On the other hand, the absence of accent, as mentioned, does not constitute evidence that an item is not an independent PrWord.

The Break Index Tier diacritics

Four diacritics are used to provide more details on the prosodic structure of utterances.

Greek (see Nespor & Vogel, 1986; Kaisse, 1985; for a different point of view, see Arvaniti, 1991, and results in Baltazani 2006b; for a review of the relevant

literature, see Arvaniti in press). Our corpus confirmed previous studies that used naturally occurring data (e.g. Fallon, 1994) in showing that sandhi can apply

across larger constituents than postulated by, e.g., Nespor & Vogel (1986); see Figures 7 and 9 for sandhi across PrWords, Figure 6 for sandhi across an ip

boundary. Further annotation of the GRToBI corpus has also shown that some of the sandhi rules of Greek are better described as gestural overlap

(Pelekanou, 2000; Arvaniti & Pelekanou 2002; Baltazani 2006b). Since the presence of sandhi does not necessarily signal cohesion, we have decided to use the

diacritic s for sandhi at all prosodic levels, and thus provide an easy way of searching the database for such instances. We believe that the sandhi phenomena

will be better understood if a large corpus of natural spoken data is investigated.

in which the context for sandhi exists but nevertheless sandhi does not take place. For example if a sequence like /ton 'pono/ ‘the pain. Acc.’ is

pronounced [ton 'pono] it should be marked as 0m, since in general it should be pronounced [to'mbono] or [to'bono]. The m diacritic should be

used with BIs 1, 2, and 3 to mark the presence of a boundary without the tonal events that normally accompany it.

The Miscellaneous Tier

This tier should be used for annotating non-structural information that may be useful in interpreting the file, such as coughing, disfluency, pitch halving or rate of speech.

Annotation examples

Figure 1: This example (‘Do the flowers really smell?’ lit. ‘the flowers smell?’), uttered in a surprised manner, illustrates the different alignment of the L*+H (the accent on ['lludia]) and L+H* (the accent on [mi'rizune]). Note the difference in the alignment of the H tone in the two accents.



 

Figure 2:  This example (‘S/he is talking to Charalampos’) shows a downstepped !H* nuclear accent. Note the lack of a dip at the beginning of this accent and compare it to the L+H* in Figure 1. 



 

3a (left)
3b (right)
Figure 3: These two examples, (both glossed as ‘they smell’), illustrate the difference between H*+L (on the left) and H* (on the right) in a one word utterance. 

 


 

Figure 4: This illustration shows two typical L* accents on segmentally identical questions, but with focus on the word [mi'rizune] ‘they smell’ on the left and on the word [lu'ludia] ‘flowers’ on the right. The questions mean ‘Do the flowers SMELL?’ and ‘Is it the FLOWERS that smell?’ respectively. Note also the different alignment of the H-, which is on (unstressed) [zu] in the question on the left, but the stressed syllable [ri] in the question on the right. 


 

Figure 5: This example (gloss: ‘Dalida was scolding the baby [when she fainted]!’) exemplifies the L-H% phrasal configuration, which is preceded in this case by a L+!H* accent on [mo'ro]. Note also the undershot (wL*+H) and early aligned (>L*+H) realizations of the two L*+H accents on ['malone] and [iDali'Da] respectively. 


 

Figure 6: This example, (‘The north wind and the sun agreed…’), illustrates the difference in scaling between H- and H-H%. Further the example shows the diphthongal realization of  /o 'iLos/, which together with the conjuction /ke/ ‘and’ form one PrWord with a rising diphthong [oI], ['coILos]; note the alignment of the L* which aligns with the whole syllable [coI]. Finally, the canonical alignment of the L*s of ['coILos] and [si'mfonisan], which are manifested as low plateaus, can be contrasted with the wL* of [mo'ro] in Figure 9. 


 

Figure 7: This example ([ta.meseona] ‘We do not live in the Middle Ages’) illustrates the typical pattern of a negative declarative expressing reservation. Note that the negative particle /Den/, which is considered a phonological clitic, carries the nuclear (and only) pitch accent of the utterance, and thus forms a separate pword from the de-accented verb ['zume] ‘we live’; yet, sandhi (/n/-deletion before the fricative [z]) does take place as well. The rest of the utterance is deaccented, with the L- spreading until after the last stressed syllable ([se] of [me'seona] ‘Middle Ages’). Finally, compare the scaling of the !H% (relative to that of the L+H* peak) to the scaling of the H- and H% tones in Figures 5, 6 and 9 relatively to the accentual Hs in those examples. 


 

 

 

 

 

 

Figure 8: This example (gloss: ‘our focus is…’) illustrates the stylized H-!H% configuration on the word ['ine] ‘is’. Note also the presence of two accents on the word [epi'cedrosi] ‘focus’, which here is followed by the enclitic [mas] ‘ours’, and thus carries enclitic stress on its last syllable [si]. 

 

 

 

Figure 9: This example, (gloss: ‘Dalida was scolding the baby when the phone rang’), shows two different realizations of L*+H under tonal crowding, >L*+H which is realized earlier than it canonically would (the H tone is aligned with the accented vowel, instead of the first postaccentual vowel), and wL*+H, in which the L* tone is undershot, while the H shows the typical late alignment of H in L*+H accents. In this utterance there is also an undershot L* (wL*) on [mo'ro], realized as a rise from low pitch throughout the accented syllable (cf. the canonical L*s in Figure 6).


 

 

Figure 10: This example, (gloss: ‘[You] BECOME-PART of society through dance’) illustrates de-accenting after early focus. Note also, the several instances of sandhi and fast speech rules. 


 


Acknowledgments
The development of GRToBI largely took place while the first author was a visitor at the Ohio State University Linguistics Laboratory. We would like to thank

the members of the Laboratory, particularly Mary Beckman, Julie McGory, Shu-hui Peng, Amanda Miller and Mariapaola D’Imperio for their encouragement and

input during the development of GRToBI, and also for long distance technical support. Thanks are also due to Georgios Tserdanelis for providing wav files for this

site, and to the students in Prof. Beckman’s and Dr McGory’s ToBI course for useful feedback at a first presentation of GRToBI. We are grateful to Sun-Ah Jun

who brought us together, suggested we develop the system and played devil's advocate at the early stages of its gestation. Finally, we would like to express our

thanks to Jenny and Peter Ladefoged for their kind hospitality to the first author during her stay at Los Angeles.

Bibliography

      II: Gesture, Segment, Prosody, 398-423. Cambridge University Press.

      University of Salzburg.

Typology: The Phonology of Intonation and Phrasing, pp. 84-117. Oxford: Oxford University Press.

      Congress of Phonetic Sciences, 4: 220-23. Stockholm.

      Laboratory Phonology V. Cambridge University Press. pp.119-131.

·         Arvaniti, A., D. R. Ladd & I. Mennen (2006a). Phonetic effects of focus and “tonal crowding” in intonation: Evidence from Greek polar questions.

      questions. Speech Communication 48: 667-696.

·         Arvaniti, A., D. R. Ladd & I. Mennen (2006b) Tonal association and tonal alignment: evidence from Greek polar questions and emphatic statements.

      statements. Language and Speech 49: 421-450.

      Actes du 5e Colloque international de linguistique grecque, pp. 71-74. Paris : L’Harmattan. 

·         Baltazani, M. (2002) Quantifier scope and the role of intonation in Greek.  Ph.D. dissertation, UCLA.

·         Baltazani, M. (2003) Broad Focus across sentence types in Greek. Proceedings of Eurospeech-2003, Geneva, Switzerland.

·         Baltazani, M. (2004) The prosodic structure of quantificational sentences in Greek. Proceedings of the 38th meeting of the Chicago Linguistic Society.

·         Baltazani, M.  (2006a) On /s/-voicing in Greek.  Proceedings of the 7th International Conference on Greek Linguistics, York, UK.

·         Baltazani, M.  (2006b) Focusing, prosodic phrasing and Hiatus resolution in Greek. In Laboratory Phonology 8, Luis Goldstein, Douglas Whalen, Catherine Best (eds.),

Mouton de Gruyter, Berlin, 473-494.

·         Baltazani, M.  (2006c) Intonation and pragmatic interpretation of negation in Greek. Journal of Pragmatics, 38, Issue 10, p. 1658-1676, Elsevier.

·         Baltazani, M. (2007a) Prosodic rhythm and the status of vowel reduction in Greek.  In Selected Papers on Theoretical and Applied Linguistics from the 17th International

Symposium on Theoretical & Applied Linguistics, Volume 1, Department of Theoretical and Applied Linguistics, Salonica, p. 31-43.

·         Baltazani, M. (2007b) Intonation of polar questions and the location of nuclear stress in Greek.  In Tones and Tunes,Volume II, Experimental Studies in Word and Sentence Prosody,

Carlos Gussenhoven & Tomas Riad (eds.), Mouton de Gruyter, Berlin, p. 387-405.

·         Baltazani, M. & Jun S. (1999) Focus and topic intonation in Greek. Proceedings of the XIVth International Congress of Phonetic Sciences, 2: 1305-8. San Fransisco.

      Prosodic Typology: The Phonology of Intonation and Phrasing, pp. 9-54. Oxford: Oxford University Press.

      Themes in Greek Linguistics, 217-224. London: John Benjamins Publishing Co.

Spoken Language Processing, 867-879.

Appendix I: Phonetic transcription conventions


In addition the following symbols and conventions should be used:

·         Noticeably centralized vowels should be transcribed as @.

·         Noticeably nasalized vowels should be transcribed with a following ~; e.g. a~ .

·         In cases of vowel coalescence, both vowels should be transcribed and joined by +; e.g. u+o resulting from a sequence of

      /u/ and /o/ (usually across a word boundary).

·         Whispered vowels should be transcribed in brackets.

·         Vowels that phonologically form separate syllables but are phonetically manifested as a rising diphthong (on the basis,

      e.g., of tonal alignment evidence), should be transcribed with the second vowel capitalized; stress should be placed before the diphthong.

·         Stress should be marked before the consonant(s) of the stressed syllable, following IPA conventions. (At present we are

      agnostic as to syllabification, so we suggest that transcribers mark maximal onsets, unless tonal alignment or their own intuitions suggests otherwise.)

Appendix II: The romanization conventions

separated by fullstops; e.g. a.i.d’oni "nightingale".

mark in the Words tier.

Home    At a Glance    Research    Publications    CV    Speech Lab    Linguistics    UCSD