The base tag set for transcriptions of spoken language described in
this chapter is intended for use with a wide variety of transcribed
spoken material. It should be stressed, however, that the present
proposals are not intended to support unmodified every variety of
research undertaken upon spoken material now or in the future; some
discourse analysts, some phonologists, and doubtless others may wish
to extend the scheme presented here to express more precisely the set
of distinctions they wish to draw in their transcriptions. Speech
regarded as a purely acoustic phenomenon may well require different
methods from those outlined here, as may speech regarded solely as a
process of social interaction.
This chapter begins with a discussion of some of the problems
commonly encountered in transcribing spoken language (section ). Section describes the basic
structural elements of this tag set. Finally, section of this chapter reviews further problems specific to the
encoding of spoken language, demonstrating how mechanisms and elements
discussed elsewhere in these Guidelines may be applied to them.
The overall structure of a TEI spoken text is identical to that of
any other TEI text: the TEI.2 element for a spoken text
contains a teiHeader element, followed by a text
element. Even texts primarily composed of transcribed speech may also
include conventional front and back matter, and may even be organized
into divisions like printed texts. For simplicity's sake, therefore, the
base tag set for spoken text uses the default text structure, as defined
in chapter ; this tag set is embedded automatically by
the spoken base tag set.
To enable the base tag set for spoken texts, a parameter entity
TEI.spoken must be declared within the document
type declaration subset, the value of which is INCLUDE, as
further described in section . A document using this
base tag set and no additional tag sets will thus begin as follows:
]>
]]>
This declaration makes available all of the elements and attributes
discussed in the present chapter, in addition to the core elements
described in chapter . If other elements are needed (in
particular, those needed for synchronization or segmentation),
additional tag sets may also be enabled in a similar way.
Two additional classes are defined by this tag set. Elements which
appear only within transcribed speech constitute the comp.spoken element class. Elements with a
specificable temporal duration constitute the timed element class. These classes are defined in
the file teispok2.ent using the following
parameter entities:
]]>
The elements of the base tag set for transcribed speech are declared in
the file teispok2.dtd, which is organized as
follows:
%TEI.structure.dtd;
]]&nil;>
]]>
General Considerations and Overview
There is great variation in the ways different researchers have
chosen to represent speech using the written medium.For a
discussion of several of these see J. A. Edwards and M. D.
Lampert, eds., Talking Language: Transcription and Coding of
Spoken Discourse (Hillsdale, N.J.: Lawrence Erlbaum Associates,
1993); Stig Johansson, Encoding a Corpus
in Machine-Readable Form, in Computational Approaches
to the Lexicon: An Overview, ed. B. T. S. Atkins et al.
(Oxford: Oxford University Press, forthcoming); and
Stig Johansson et al. Working Paper on Spoken
Texts, document TEI AI2 W1, 1991. This
reflects the special difficulties which apply to the encoding or transcription of speech. Speech varies according to
a large number of dimensions, many of which have no counterpart in
writing (for example, tempo, loudness, pitch, etc.). The audibility of
speech recorded in natural communication situations is often less than
perfect, affecting the accuracy of the transcription. Spoken material
may be transcribed in the course of linguistic, acoustic,
anthropological, psychological, ethnographic, journalistic, or many
other types of research. Even in the same field, the interests and
theoretical perspectives of different transcribers may lead them to
prefer different levels of detail in the transcript and different styles
of visual display. The production and comprehension of speech are
intimately bound up with the situation in which speech occurs, far more
so than is the case for written texts. A speech transcript must
therefore include some contextual features; determining which are
relevant is not always simple. Moreover, the ethical problems in
recording and making public what was produced in a private setting and
intended for a limited audience are more frequently encountered in
dealing with spoken texts than with written ones.
Speech also poses difficult structural problems. Unlike a written
text, a speech event takes place in time. Its beginning and end may be
hard to determine and its internal composition difficult to define.
Most researchers agree that the utterances or turns of
individual speakers form an important structural component in most kinds
of speech, but these are rarely as well-behaved (in the structural
sense) as paragraphs or other analogous units in written texts:
speakers frequently interrupt each other, use gestures as well as words,
leave remarks unfinished and so on. Speech itself, though it may be
represented as words, frequently contains items such as vocalized pauses
which, although only semi-lexical, have immense importance in the
analysis of spoken text. Even non-vocal elements such as gestures may
be regarded as forming a component of spoken text for some analytic
purposes.
Below the level of the individual utterance, speech may be segmented
into units defined by phonological, prosodic, or syntactic phenomena;
no clear agreement exists, however, even as to appropriate names for
such segments.
Spoken texts transcribed according to the guidelines presented here
are organized as follows. As noted above, speech is regarded as being
composed of arbitrary high-level units called
texts.
A spoken text might typically be a conversation between a small
number of people, a lecture, a broadcast TV item, or a similar event.
Each such unit has associated with it a teiHeader providing
detailed contextual information such as the source of the transcript,
the identity of the participants, whether the speech is scripted or
spontaneous, the physical and social setting in which the discourse
takes place and a range of other aspects. For details of the header in
general, refer to chapter ; for details of additional
elements for the documentation of participant and contextual
information, see section .
Defining the bounds of a spoken text is frequently a matter of
arbitrary convention or convenience. In public or semi-public contexts,
a text may be regarded as synonymous with, for example, a lecture, a broadcast item,
a meeting, etc. In informal or private
contexts, a text may be simply a conversation involving a specific group
of participants. Alternatively, researchers may elect to define spoken
texts solely in terms of their duration in time or length in words. By
default, these Guidelines assume of a text only that:
it is internally cohesive,
it is describable by a single header, and
it represents a single stretch of time with no significant
discontinuities.
Deviation from these assumptions may be specified (for example, the
org attribute on the text element may take the value
compos to specify that the components of the
text are discrete) but is not recommended.
Within a text it may be necessary to identify subdivisions
of various kinds, if only for convenience of handling. The neutral
div element discussed in section is
recommended for this purpose. It may be found useful also for
representing subdivisions relating to discourse structure, speech act
theory, transactional analysis, etc., provided that these divisions are
hierarchically well-behaved.
Where they are not,
as is often the case, the mechanisms discussed in chapters
and
may be used.
A spoken text may contain any of the following components:
utterances
pauses
vocalized but non-lexical phenomena such as coughs
kinesic (non-verbal, non-lexical) phenomena such as gestures
entirely non-linguistic events occurring during and possibly
influencing the course of speech
writing, regarded as a special class of event in that it can
be transcribed, for example captions or overheads displayed during
a lecture
shifts or changes in vocal quality
Elements to represent all of these features of spoken language are
discussed in section below.
An utterance (tagged u) may contain lexical items
interspersed with pauses and non-lexical vocal sounds; during an
utterance, non-linguistic events may occur and written materials may be
presented. The u element can thus contain any of the other
elements listed, interspersed with a transcription of the lexical items
of the utterance; the other elements may all appear between utterances
or next to each other, but except for writing they do not
contain any other elements nor any data.
Divisions
A spoken text itself may be without substructure, that is, it may
consist simply of units such as utterances or pauses, not grouped
together in any way, or it may be subdivided into one or more divisions
as described in this section.
If the notion of what constitutes a text in
spoken discourse is inevitably rather an arbitrary one, the notion of
formal subdivisions within such a text is even more
debatable. Nevertheless, such divisions may be useful for such types of
discourse as debates, broadcasts, etc., where structural subdivisions
can easily be identified, or more generally wherever it is desired to
aggregate utterances or other parts of a transcript into units smaller
than a complete text. Examples might include
conversations or discourse fragments, or more narrowly,
that part of the conversation where topic x was discussed,
provided only that the set of all such divisions is coextensive with the
text.
Each such division of a spoken text should be represented by the
numbered or un-numbered div elements defined in chapter . For some detailed kinds of analysis a hierarchy of such
divisions may be found useful; nested div elements may be used
for this purpose, as in the example below.
The div element is a member of the divn class of structural elements, and therefore has
the following attributes in common with other members of the class:
specifies a name conventionally used for this level of
subdivision, e.g. act, volume, book,
section, canto, etc.specifies how the content of the division is organized.
Legal values are:
composite content: i.e. no claim is made about the sequence
in which the immediate contents of this division are to be
processed, or their inter-relationships.uniform content: i.e. the immediate contents of this
element are regarded as forming a logical unit, to be
processed in sequence.indicates whether this division is a sample of the original
source and if so, from which part.
Legal values are:
division lacks material present at end in source.division lacks material at start and end.division lacks material at start.position of sampled material within original unknown.division is not a sample.specifies whether or not the division is fragmented by some
other structural element, for example a speech which is
divided between two or more verse stanzas.
Legal values are:
the division is incomplete in some respecteither the division is complete, or no claim is made as to
its completeness.the initial part of an incomplete divisiona medial part of an incomplete divisionthe final part of an incomplete division
The type attribute may be used to characterize divisions
in any way that is convenient; no specific recommendations are made in
these Guidelines. For example, a collection made up of transcribed
sound bites taken from speeches given by a
politician on different occasions, might encode each extract as a
distinct div, nested within a single composite div as
follows:
]]>
As a member of the class declaring, the
div element may also carry a decls attribute, for
use where the divisions of a text do not all share the same set of the
contextual declarations specified in the TEI header. (See further
section ).
Elements Unique to Spoken Texts
The following elements characterize spoken texts, transcribed
according to these Guidelines:
a stretch of speech usually preceded and followed by
silence or by a change of speaker.a pause either between or within utterances.any vocalized but not necessarily lexical phenomenon, for
example
voiced pauses, non-lexical backchannels, etc.any communicative phenomenon, not necessarily vocalized,
for
example a gesture, frown, etc.any phenomenon or occurrence, not necessarily vocalized or
communicative, for example incidental noises or other
events affecting
communication.a passage of written text revealed to participants in the
course
of a spoken text.marks the point at which some paralinguistic feature of a
series of
utterances by any one speaker changes.
Each of these is further discussed and specified below in sections
to .
We can show the relationship between four of these constituents of
speech using the features eventive, communicative, anthropophonic (for sounds produced by the human
vocal apparatus), and lexical:
The differences are not always clear-cut. Among
events might be included actions like
slamming the door, which can certainly be communicative.
Vocals include coughing and sneezing, which are
usually involuntary noises. Equally, the distinction between utterances
and vocals is not always clear,
although for many analytic purposes it will be convenient to regard them
as distinct. Individual scholars may differ in the way borderlines are
drawn and should declare their definitions in the editorialDecl
element of the header (see ).
The following short extract exemplifies several of these elements. It
is recoded from a text originally transcribed in the CHILDES
format.The original is a conversation between two children and
their parents, recorded in 1987, and discussed in
Brian MacWhinney,CHAT Manual
([Pittsburgh]: Dept of Psychology, Carnegie-Mellon University,
1988), pp. 87ff.
Each utterance is encoded using a u element (see section ). Pauses marked by the transcriber are indicated using
the pause element (see section ).
Non-verbal vocal effects such as the child's meowing are indicated
either with orthographic transcriptions or with the vocal
element, and entirely non-linguistic but significant events such as the
sound of the toy cat are represented by the event elements
(see section ).
you never take this cat for show&sp;and&sp;tell
meow meow
yeah well I dont want tobecause it is so oldhow about your&stress; cat
yours is new &stress;
thats darlingno mine&stress; isnt oldmine is just um a little dirty
]]>
This example also uses some elements common to all TEI texts,
notably the reg tag for editorial regularization. Special
purpose entity references have been used to indicate non-separating
spaces (&sp) and unusually stressed syllables
(&stress); an alternative to the latter might
have been to use the core emph element. The s element
has also been used to segment the last utterance. Further discussion of
all of these options is provided in section .
Contextual information is of particular importance in spoken texts,
and is usually provided by the TEI header of a text. In general, all
of the information in a header is understood to be relevant to the whole
of the associated text. The elements u and writing
are however members of the declaring class,
and may therefore specify a different context from that of the
surrounding elements within a given division or text by means of the
decls attribute (see further section ).
Utterances
Each distinct utterance in a spoken text is represented
by a u element, described as follows:
a stretch of speech usually preceded and followed by
silence or by a change of speaker.
Attributes include:
supplies an identifier for the speaker or group of
speakers. Its value is the identifier of a
participant or participantGrp element
in the TEI header.indicates the nature of the transition between this
utterance and the previous one.
Sample values include:
this utterance begins without unusual pause or rapidity.this utterance begins with a markedly shorter pause than
normal.this utterance begins before the previous one has finished.this utterance begins after a noticeable pause.
Use of the who attribute to associate the utterance with a
particular speaker is recommended but not required. Its use implies as
a further requirement that all speakers be identified by a
participant or participantGrp element in the TEI
header (see section ). Where utterances cannot be
attributed with confidence to any particular participant or group of
participants, the encoder may choose to define
participants such as all or
various. For example:
]]>
The trans attribute is provided as a means of
characterizing the transition from one utterance to the next at a
simpler level of detail than that provided by the temporal alignment
mechanism discussed in section . The value specified
applies to the transition from the preceding utterance into the
utterance bearing the attribute. For example:For
the most part, the examples in this chapter use no sentence punctuation
except to mark the rising intonation often found in interrogative
statements; for further discussion, see section .
Have you heard the
the election results? yesit's a disasterit's a miracle
]]>
In this example, utterance B1 latches on to utterance A1, while there is
a marked pause between B1 and A2. B2 and A2 overlap, but by an
unspecified amount. For ways of providing a more precise indication of
the degree of overlap, see section .
An utterance may contain either running text, or text within which
other basic structural elements are nested. Where such nesting occurs,
the who attribute is considered to be inherited for the
elements pause, vocal, shift and
kinesic; that is, a pause or shift (etc.) within an utterance
is regarded as being produced by that speaker only, while a pause
between utterances applies to all speakers.
Occasionally, an utterance may contain other utterances, for example
where there is a change in the script associated with it. This may
occur when a speaker changes script in mid-utterance. For example:
Listen to this
The government is confident, he
said, that the current economic problems will be
completely overcome by June
what nonsense
]]>
Here speaker A's own utterance contains a second nested utterance, which
is read from a newspaper. The decls attribute on the nested
utterance is used to indicate that its script is S1, rather than the
default. Alternatively, the embedded utterance might be regarded as a
new (non-nested) one. It might also be encoded using the
writing element described in section
below, or the event element described in section , without transcribing the read material:
Listen to this
what nonsense
]]>
Pause
The pause empty element is used to indicate a perceived
pause, either between or within utterances.
a pause either between or within utterances.
Attributes include:
supplies an identifier for the person or group pausing. Its
value is the identifier of a participant or
participant.grp element in the TEI header.categorizes the pause in some respect.
A pause contained by an utterance applies to the speaker of that
utterance. A pause between utterances applies to all speakers. The
type attribute may be used to categorize the pause, for
example as short, medium or long; alternatively the attribute
dur may be used to indicate its length more exactly, as in
the following example:
Okay U-mthe s the scene
opens up with um
you see a tree okay?
]]>
If detailed synchronization of pausing with other vocal phenomena is
required, the alignment mechanism defined at section
and discussed informally below should be used. Note that the
trans attribute mentioned in the previous section may also be
used to characterize the degree of pausing between (but not within)
utterances.
Vocal, Kinesic, Event
These three empty elements are used to indicate the presence of
non-transcribed semi-lexical or non-lexical phenomena either between or
within utterances.
any vocalized but not necessarily lexical phenomenon, for
example
voiced pauses, non-lexical backchannels, etc.
Attributes include:
supplies an identifier for the vocalist(s). Its value is
the identifier of a participant or
participant.grp element in the TEI header.supplies a conventional representation for the phenomenon.indicates whether or not the phenomenon is repeated.
Sample values include:
the phenomenon is repeated.the phenomenon is atomic.unknown or unmarked.any communicative phenomenon, not necessarily vocalized,
for
example a gesture, frown, etc.
Attributes include:
supplies an identifier for the participant performing the
gesture. Its value is the identifier of a
participant or participant.grp
element in the TEI header.supplies a conventional representation for the phenomenon.indicates whether or not the phenomenon is repeated.
Sample values include:
the phenomenon is repeated.the phenomenon is atomic.unknown or unmarked.any phenomenon or occurrence, not necessarily vocalized or
communicative, for example incidental noises or other
events affecting
communication.
Attributes include:
supplies an identifier for the agent of the event
described, if any. Its value is the identifier of a
participant or participant.grp
element in the TEI header.supplies a conventional representation for the phenomenon.indicates whether or not the phenomenon is repeated.
Sample values include:
the phenomenon is repeated.the phenomenon is atomic.unknown or unmarked.
The who attribute should be used to specify the person or
group responsible for a vocal, kinesic or event which is contained
within an utterance, if this differs from that of the enclosing
utterance. The attribute must be supplied for a vocal, kinesic or event
which is not contained within an utterance.
The iterated attribute may be used to indicate that the
vocal, kinesic or event is repeated, for example laughter as opposed to laugh.
These should both be distinguished from laughing,
where what is being encoded is a shift in voice quality. For this last
case, the shift element discussed in section should be used.
The desc attribute may be used to supply a conventional
representation for the phenomenon, for example:
Researchers may prefer to regard some semi-lexical phenomena as
words within the bounds of the u element.
See further the discussion at section below. As
for all basic categories, the definition should be made clear in the
encodingDesc element of the TEI header.
Some typical examples follow (recoded from
J. Maxwell Atkinson and John Heritage, eds.,Transcript notation. Structures of social action: Studies in
conversation analysis, Cambridge University Press,
1984).
This is just delicious
I'll get itI used to smoke a lotHe thinks he's tough
]]>
Note that Ann's snorting could equally well be encoded as follows:
]]>
The extent to which encoding of events or kinesics is included in a
transcription will depend entirely on the purpose for which the
transcription was made. As elsewhere, this will depend on the
particular research agenda and the extent to which their presence is
felt to be significant for the interpretation of spoken interactions.
Writing
Written text may also be encountered when speech is transcribed, for
example in a television broadcast or cinema performance, or where one
participant shows written text to another. The writing element
may be used to distinguish such written elements from the spoken text in
which they are embedded.
writinga passage of written text revealed to participants in the
course
of a spoken text.
Attributes include:
whosupplies an identifier for the participant who reveals or
creates the writing, if any. Its value is the identifier of
a participant or participant.grp
element in the TEI header.gradualindicates whether the writing is revealed all at once or
gradually.
Sample values include:
ythe writing is revealed gradually.nthe writing is revealed all at once.uunknown or unmarked.typecategorizes the kind of writing in some way, for example as
a subtitle, noticeboard etc.
For example, if speaker A in the breakfast table conversation in section
above had simply shown the newspaper passage to her
interlocutor instead of reading it, the interaction might have been
encoded as follows:
look at this
The government is confident, he said, that the
current economic problems will be completely overcome
by June
what nonsense!
]]>
Temporal Information
In addition to the global attributes n, id, and
lang, utterances, vocals, pauses, kinesics, events and
writing elements may all take a common set of attributes providing
information about their position in time. For this reason, these
elements are regarded as forming a class, referred to here
as timed. The following attributes are common
to all elements in this class:
startindicates the location within a temporal alignment at which
this element begins.endindicates the location within a temporal alignment at which
this element ends.durindicates the length of this element in time, using either
specific units or the units specified on the associated
temporal alignment.
Note that if start and end point to when
elements whose temporal distance from each other is specified in a
timeline, then dur is ignored.
The anchor element (see ) may be used as
an alternative means of aligning the start and end of timed elements,
and is required when the temporal alignment involves points within an
element.
Shifts
A common requirement in transcribing spoken language is to mark
positions at which a variety of prosodic features change. Many
paralinguistic features (pitch, prominence, loudness, etc.) characterize
stretches of speech which are not co-extensive with utterances or any of
the other units discussed so far. One simple method of encoding such
units is simply to mark their boundaries. An empty element called
shift is provided for this purpose.
shiftmarks the point at which some paralinguistic feature of a
series of
utterances by any one speaker changes.
Attributes include:
featurea paralinguistic feature.
Sample values include:
tempospeed of utterance.loudloudness.pitchpitch range.tensiontension or stress pattern.rhythmrhythmic qualities.voicevoice quality.newspecifies the new state of the paralinguistic feature
specified.
A shift element may appear within an utterance or a segment to
mark a significant change in the particular feature defined by its
attributes, which is then understood to apply to all subsequent
utterances for the same speaker, unless changed by a new shift for the
same feature in the same speaker. Intervening utterances by other
speakers do not normally carry the same feature.
For example:
Elizabeth
YesCome and try this come on
]]>
In this example, the word Elizabeth is spoken
loudly, the words Yes and Come and try
this with normal volume, and the words come
on very loudly.
The values proposed here for the feature attribute are
based on those used by the Survey of English Usage;
For details see S. Boase, London-Lund
Corpus: Example Text and Transcription Guide (London:
Survey of English Usage, University College London,
1990).
this list may be revised or supplemented using the methods outlined in
section .
The new attribute specifies the new state of the feature
following the shift. If no value is specified, it is implied that the
feature concerned ceases to be remarkable at this point: the special
value normal may be specified to have the same
effect.
A list of suggested values for each of the features proposed follows:
tempo
a allegro (fast)aa very fastacc accelerando (getting faster)l lento (slow)ll very slowrall rallentando (getting slower)loud (for loudness):
f forte (loud)ff very loudcresc crescendo (getting louder)p piano (soft)pp very softdimin diminuendo (getting softer)pitch (for pitch range):
high high pitch-rangelow low pitch-rangewide wide pitch-rangenarrownarrow pitch-rangeasc ascendingdesc descendingmonot monotonousscand scandent, each succeeding syllable higher than
the last, generally ending in a falling tonetension:
sl slurredlax lax, a little slurredten tensepr very precisest staccato, every stressed syllable being doubly
stressedleg legato, every syllable receiving more or less equal
stressrhythm:
rh beatable rhythmarrh arrhythmic, particularly haltingspr spiky rising, with markedly higher unstressed
syllablesspf spiky falling, with markedly lower unstressed
syllablesglr glissando rising, like spiky rising but the
unstressed syllables, usually several, also rise
in pitch relative to each otherglf glissando falling, like spiky falling but with the
unstressed syllables also falling in pitch relative
to each othervoice (for voice quality):
whisp whisperbreath breathyhusk huskycreak creakyfals falsettoreson resonantgiggle unvoiced laugh or gigglelaugh voiced laughtrem tremuloussob sobbingyawn yawningsigh sighing
A full definition of the sense of the values provided for each
feature should be provided in the encoding description section of the
text header (see section ).
Formal Definition
The components of the tag set for transcribed speech are formally
defined as follows:
]]>
Elements Defined Elsewhere
This section describes the following features characteristic of
spoken texts for which elements are defined elsewhere in these
Guidelines:
segmentation below the utterance level
synchronization and overlap
regularization of orthography
The elements discussed here are not provided by the base tag set for
spoken texts. Some of them are included in the core tag set available
to all TEI documents, but others are contained in the TEI additional
tag sets for linking and for analysis respectively. To enable these tag
sets, the appropriate parameter entities must be declared in the
document type declaration subset, as described in section . For example, if a transcript using the base tag set
defined in this chapter additionally wishes to make use of the
timeLine element, then the following declarations would be
necessary:
]>
]]>
If the complex segmentation elements defined in the additional tag set
for analysis were also required, the following declarations would be
needed:
]>
]]>
Segmentation
For some analytic purposes it may be desirable to subdivide the
divisions of a spoken text into units smaller than the individual
utterance or turn. Segmentation may be performed for a number of
different purposes and in terms of a variety of speech phenomena.
Common examples include units defined both prosodically (by intonation,
pausing, etc.) and syntactically (clauses, phrases, etc.) The term
macrosyntagm has been used by a number of researchers to
define units peculiar to speech transcripts. The term was
apparently first proposed by Bengt Loman and Nils
Jørgensen, in Manual for analys och beskrivning
av makrosyntagmer (Lund: Studentlitteratur, 1971),
where it is defined as follows: A text can be analysed as a sequence
of segments which are internally connected by a network of syntactic
relations and externally delimited by the absence of such relations with
respect to neighbouring segments. Such a segment is a syntactic unit
called a macrosyntagm (trans. S. Johansson).
These Guidelines propose that such analyses be performed in terms of
neutrally-named segments, represented by the seg
element, which is discussed more fully in section .
This element may take a type attribute to specify the kind of
segmentation applicable to a particular segment, if more than one is
possible in a text. A full definition of the segmentation scheme or
schemes used should be provided in the segmentation element of
the editorialDecl element in the TEI header (see ).
In the first example below, an utterance has been segmented according
to a notion of syntactic completeness not necessarily marked by the
speech, although in this case a pause has been recorded between the two
sentence-like units. In the second, the segments are defined
prosodically (an entity reference &stress;
has been used to mark the position immediately following the syllable
bearing the primary accent or stress), and may be thought of as
tone units.
we went to the pub yesterdaythere was no one therealthough its an old ide&stress;ait hasnt been on the mar&stress;ket very long
]]>
In either case, the segmentation element in the header of the
text should specify the principles adopted to define the segments marked
in this way.
When utterances are segmented end-to-end in the same way as the
s-units in written texts, the s element discussed in chapter
may be used, either as an alternative or in addition to
the more general purpose seg element. The s element
is available without formality in all texts, but does not allow segments
to nest within each other.
Where segments of different kinds are to be distinguished within the
same stretch of speech, the type attribute may be used, as in
the following example. The example also shows the use of a
user-specified extension to the TEI tag sets, for specifying
paraphasia.
I think this chap was writing and he said hello said hello and he said I'm going to a gate
at twenty past seven he said ok right away and so on they went and they were
writing there
]]>
In this example, recoded from a corpus of language-impaired speech
prepared by Fletcher and Garman, the speaker's utterance has been fully
segmented into clausal (type=C) or minor (type=M) units. An additional element paraphasia has been used to define a particular
characteristic of this corpus for which no element exists in the TEI
scheme.
See further chapter for a discussion of the way in
which this kind of user-defined extension of the TEI scheme may be
performed and chapter for the mechanisms on which it
depends.
This example also uses the core elements omit and
del to mark editorial decisions concerning matter completely
omitted from the transcript (because of inaudibility), and words which
have been transcribed but which the transcriber considers may be
deleted, respectively. See further section for a
discussion of these and related elements.
It is often the case that the desired segmentation does not respect
utterance boundaries; for example, syntactic units may cross utterance
boundaries. For a detailed discussion of this problem, and the various
methods proposed by these Guidelines for handling it, see chapter . Methods discussed there include these:
a concurrent DTD may be defined
milestone tags may be used;
the special-purpose shift tag discussed
in section is an extension of this method
where several discontinuous segments are to be grouped
together to form a syntactic unit (e.g. a phrasal verb with interposed
complement), the join element may be used
Synchronization and Overlap
A major difference between spoken and written texts is the importance
of the temporal dimension to the former. As a very simple example,
consider the following, first as it might be represented in a
playscript:
Let us assume that Stig and Lou respond to Jane's question before she
has finished asking it --- a fairly normal situation in spontaneous
speech. The simplest way of representing this overlap
would be to use the trans attribute previously discussed:
have you read Vanity Fair
yes
]]>
However, this does not allow us to indicate either the extent to which
Jane's utterance is overlapped, nor does it show that there are in
fact three things which are synchronous: the end of Jane's utterance,
Stig's whole utterance, and Lou's kinesic. To overcome these problems,
more sophisticated techniques, employing the mechanisms for pointing and
alignment discussed in detail in section , are needed.
If the additional tag set for linking has been enabled (as described in
section above), one way to represent the simple
example above would be as follows:
have you read Vanity
Fair
yes
]]>
For a full discussion of this and related mechanisms, section should be consulted. The rest of the present section,
which should be read in conjunction with that more detailed discussion,
presents a number of ways in which these mechanisms may be applied to
the specific problem of representing temporal alignment, synchrony or
overlap in transcribing spoken texts.
In the simple example above, the first utterance (that with
identifier u1) contains an anchor element, the function of
which is simply to mark a point within it. The synch
attribute associated with this anchor point specifies the identifiers of
the other two elements which are to be synchronized with it:
specifically, the second utterance (u2) and the kinesic (k1). Note that
one of these elements has content and the other is empty.
This example demonstrates only a way of indicating a point within one
utterance at which it can be synchronized with another utterance and a
kinesic. For more complex kinds of alignment, involving possibly
multiple synchronization points, an additional element is provided,
known as a timeLine. This consists of a series of
when elements, each representing a point in time, and bearing
attributes which indicate its exact temporal position relative to other
elements in the same timeline, in addition to the sequencing implied by
its position within it.
For example:
]]>
This timeline represents four points in time, named P1, P2, P6, and P3
(as with all attributes named id in the TEI scheme, the
names must be unique within the document but have no other
significance). P1 is located absolutely, at 12:20:01:01 BST. P2 is 4.5
seconds (i.e. 45 deci-seconds) later than P2 (i.e. at 12:20:46). P6 is
at some unspecified time later than P2 and previous to P3 (this is
implied by its position within the timeline, as no attribute values have
been specified for it). The fourth point, P3, is 1.5 seconds (15 dsec)
later than P6.
One or more such timelines may be specified within a spoken text, to
suit the encoder's convenience. If more than one is supplied, the
origin attribute may be used on each to specify which other
timeLine element it follows. The unit attribute
indicates the units used for timings given on when elements
contained by the alignment map. Alternatively, to avoid the need to
specify times explicitly, the interval attribute may be used
to indicate that all the when elements in a time line are a
fixed distance apart.
Three methods are available for aligning points or elements within a
spoken text with the points in time defined by the timeLine:
The elements to be synchronized may specify the identifier
of a when element as the value of one of the start,
end or synch attributes The when
element may specify the identifiers of all the elements to be
synchronized with it using the synch attribute A
free-standing link element may be used to associate the
when element and the elements synchronized with it by
specifying their identifiers as values for its target
attribute.
For example, using the timeline given above:
This is my turn
]]>
The start of this utterance is aligned with P2 and its end with P3. The
transition between the words my and
turn occurs at point P6.
The synchronization represented by the preceding examples could
equally well be represented as follows:
...
This is my turn
]]>
Here, the whole of the object with identifier U1 (the utterance) has
been aligned with two different points, P2 and P3. This is interpreted
to mean that the utterance spans at least those two points.
Finally, a linkGrp may be used as an alternative to the
synch attribute:
...
This is my turn
...
]]>
As a further example of the three possibilities, consider the
following dialogue, represented first as it might appear in a
conventional playscript:
A commonly used convention might be to transcribe such a passage as
follows:
I used to smoke [ a lot more than this ]
<2> [ you used to smoke ]
<1> but I never inhaled the smoke
]]>
Such conventions have the drawback that they are hard to generalize or
to extend beyond the very simple case presented here. Their reliance on
the accidentals of physical layout may also make them difficult to
transport and to process computationally. These Guidelines recommend
one of the courses described in what follows:
Where the whole of one or another utterance is to be synchronized,
the start and end attributes may be used:
I used to smoke a lot more than this
but I never inhaled the smoke
You used to smoke
]]>
Note that the second utterance above could equally well be encoded as
follows with exactly the same effect:
You used to smoke
]]>
If synchronization with specific timing information is required, a
timeLine must be included:
....
I used to smoke
a lot more than this
but I never inhaled the smokeYou used to smoke
]]>
As above, since the whole of Bob's utterance is to be aligned, the
start and end attributes may be used as an
alternative to the second pair of anchor elements:
You used to smoke
]]>
An alternative approach is to mark the synchronization by pointing
from the timeLine to the text:
....
I used to smoke
a lot more than this
but I never inhaled the smokeYou used to smoke
]]>
To avoid deciding whether to point from the timeline to the text or vice
versa, a linkGrp may be used:
I used to smoke
a lot more than this
but I never inhaled the smokeYou used to smoke
!--- ... -->
]]>
Note that in each case, although Bob's utterance follows Tom's
sequentially in the text, it is aligned temporally with its middle,
without any need to disrupt the normal syntax of the text.
As a final example, consider the following exchange, first as it
might be represented using a musical-score-like notation, in which
points of synchronization are represented by vertical alignment of the
text:
All three speakers are simultaneous at the words
my, Balderdash, and
No; speakers A and C are simultaneous at the
words turn and it's. This
could be encoded as follows, using pointers from the alignment map
into the text:
...
this is my turnbalderdash no it's mine
]]>
Regularization of Word Forms
When speech is transcribed using ordinary orthographic notation, as
is customary, some compromise must be made between the sounds produced
and conventional orthography. Particularly when dealing with informal,
dialectal or other varieties of language, the transcriber will
frequently have to decide whether a particular sound is to be treated as
a distinct vocabulary item or not. For example, while in a given
project kinda may not be worth distinguishing as a
vocabulary item from kind of,
isn't may clearly be worth distinguishing from
is not; for some purposes, the regional variant
isnae might also be worth distinguishing in the
same way.
One rule of thumb might be to allow such variation only where a
generally accepted orthographic form exists, for example, in published
dictionaries of the language register being encoded; this has the
disadvantage that such dictionaries may not exist. Another is to
maintain a controlled (but extensible) set of normalized forms for all
such words; this has the advantage of enforcing some degree of
consistency among different transcribers. Occasionally, as for example
when transcribing abbreviations or acronyms, it may be felt necessary to
depart from conventional spelling to distinguish between cases where the
abbreviation is spelled out letter by letter (e.g. B B
C or V A T) and where it is
pronounced as a single word (VAT or
RADA). Similar considerations might apply to
pronunciation of foreign words (e.g. Monsewer vs.
Monsieur).
In general, use of punctuation, capitalization, etc., in spoken
transcripts should be carefully controlled. It is important to
distinguish the transcriber's intuition as to what the punctuation
should be from the marking of prosodic features such as pausing,
intonation, etc.
Whatever practice is adopted, it is essential that it be clearly and
fully documented in the editorial declarations section of the header.
It may also be found helpful to include normalized forms of
non-conventional spellings within the text, using the elements for
simple editorial changes described in section (see
further section ).
Prosody
In the absence of conventional punctuation, the marking of prosodic
features assumes paramount importance, since these structure and
organize the spoken message. Indeed, such prosodic features as points
of primary or secondary stress may be represented by specialized
punctuation marks. Pauses have already been dealt with in section ; while tone units (or intonational phrases) can be
indicated by the segmentation tag discussed in section . The shift tag discussed in section
may also be used to encode some prosodic features, for example where all
that is required is the ability to record shifts in voice quality.
For more detailed work, involving a detailed phonological transcript
including representation of stress and pitch patterns, it is probably
best to maintain the prosodic description in parallel with the
conventional written transcript, rather than attempt to embed detailed
prosodic information within it. The two parallel streams may be aligned
with each other and with other streams, for example an acoustic
encoding, using the general alignment mechanisms discussed in section
.
Where only a small number of phonetic or phonemic aspects are
included in a transcript, it may be convenient to provide a simple set
of entity declarations for the particular set of features marked. The
entity references in the text may then be redefined to produce simple
punctuation marks (as in the following example), or as references to
bundles of phonological features, in the same way as is proposed for
part of speech tags (see section ).
In the following example, a small set of prosodic features are
recorded throughout the transcript using a user-defined entity set such
as the following:
]]>
This set of entity definitions may be included directly within the
document type declaration subset for the file, or more conveniently
along with any other extensions or modifications within the user
extensions file defined by the entity TEI.extensions.ent, as discussed in section . For convenience of reading on the screen, these entity
declarations will map the mnemonic entity names used in the text below
to a conventional punctuation mark.
C is with a friendExcuse me&lf;
You dont have some aesthetic&trunc;
specially on early
aesthetics terminology &lr;No&lf; No&lf;
I'm afraid&lf;No&lr;
Well thanks&lr;
Oh&trunc; you couldnt&trunc can we
kind of&long; I mean ask you to order
it for us&long;&fr;Yes&fr;
if you know the title&lf; Yeah&lf;Yes thats fine.
just as soon as it comes in we'll send you
a postcard&lf
]]>
This example, which is taken from a corpus of bookshop service
encountersLaura Gavioli and Gillian
Mansfield, The Pixi Corpora (Bologna:
Cooperativa Libraria Universitaria Editrice, 1990), p. 74.
also demonstrates the use of the unclear and omit
elements discussed in section . Where words are so
unclear that only their extent can be recorded, the empty omit
element may be used; where the encoder can identify the words but wishes
to record a degree of uncertainty about their accuracy, the
unclear element may be used. More flexible and detailed
methods of indicating uncertainty are discussed in chapter .
Where a transcript includes many phonetic or phonemic aspects, it
will generally be convenient to use a specialized writing system, as
defined in chapters and . For
representation of phonemic information, the use of the International
Phonetic Alphabet is recommended.
Speech Management
Phenomena of speech management include disfluencies such
as filled and unfilled pauses, interrupted or repeated words,
corrections, and reformulations as well as interactional devices asking
for or providing feedback. Depending on the importance attached to such
features, transcribers may choose to adopt conventionalized
representations for them (as discussed in section
above), or to transcribe them using IPA or some other transcription
system. To simplify analysis of the lexical features of a speech
transcript, it may be felt useful to tidy away many
of these disfluencies. Where this policy has been adopted, these
Guidelines recommend the use of the tags for simple editorial
intervention discussed in section , to make explicit
the extent of regularization or normalization performed by the
transcriber.
For example, false starts, repetition, and truncated words might all
be included within a transcript, but marked as editorially deleted, in
the following way:
ssee
you youyou know
it'she's crazy
]]>
As previously noted, the gap element may be used to mark
points within a transcript where words have been omitted, for example
because they are inaudible:
]]>
The unclear element may be used to mark words which have
been included although the transcriber is unsure of their accuracy:
marbled queen
]]>
Where a transcriber is believed to have incorrectly identified a
word, the elements corr or sic may be used to indicate
both the original and a corrected form of it:
skuzzy
SCSI
]]>
As discussed in section , the first of these would
be appropriate where faithfulness to the transcribers' intuition is
paramount, and the second where the editorial interpretation is felt
more significant. In either case, the user of the text can perceive the
basis of the choice being offered.
Analytic Coding
The recommendations made here only concern the establishment of a
basic text. Where a more sophisticated analysis is needed, more
sophisticated methods of markup will also be appropriate, for example,
using concurrent markup streams to indicate multiple segmentation of the
stream of discourse, or complex alignment of several segments within it.
Where additional annotations (sometimes called
codes or tags) are used to
represent such features as linguistic word class (noun, verb, etc.),
type of speech act (imperative, concessive, etc.), or information status
(theme/rheme, given/new, active/semi-active/new), etc., a selection from
the general purpose analytic tools discussed in chapters , , and , may be used to
advantage.