PROJECTING THE END OF A SPEAKER’S TURN:
A COGNITIVE CORNERSTONE OF CONVERSATION
N. J. E
Max Planck Institute for Max Planck Institute for Max Planck Institute for
Psycholinguistics Psycholinguistics Psycholinguistics
A key mechanism in the organization of turns at talk in conversation is the ability to anticipate
the moment of completion of a current speaker’s turn. Some authors suggest that this
is achieved via lexicosyntactic cues, while others argue that projection is based on intonational
contours. We tested these hypotheses in an on-line experiment, manipulating the presence of
symbolic (lexicosyntactic) content and intonational contour of utterances recorded in natural con-
versations. When hearing the original recordings, subjects can anticipate turn endings with the
same degree of accuracy attested in real conversation. With intonational contour entirely removed
(leaving intact words and syntax, with a completely flat pitch), there is no change in subjects’
accuracy of end-of-turn projection. But in the opposite case (with original intonational contour
intact, but with no recognizable words), subjects’ performance deteriorates significantly. These
results establish that the symbolic (i.e. lexicosyntactic) content of an utterance is necessary (and
possibly sufficient) for projecting the moment of its completion, and thus for regulating conversa-
tional turn-taking. By contrast, and perhaps surprisingly, intonational contour is neither necessary
nor sufficient for end-of-turn projection.*
. Getting one’s timing right is a key problem in speaking. When
producing and comprehending speech in conversation, we come under a range of psy-
chological and performance pressures, requiring both speed and temporal accuracy. In
the flow of interaction, we run a battery of simultaneous tasks: we are perceiving and
processing the speech of others; we are formulating our own utterances in advance;
we are simultaneously monitoring the internal timing of our own speech and the timing
of our own utterances relative to those of our interlocutors; we are monitoring the
content of our own speech and correcting problems if detected; we are monitoring
others’ responses to our utterances and correcting problems if detected; we are produc-
ing and comprehending hand gestures and other bodily actions linked to the speech;
and much more besides. Among this rich and urgent flow of perceptual information
and motor activity, not only do we work to produce utterances that are well-formed
and that achieve the purposes they are designed to achieve (e.g. eliciting information,
prompting action, etc.), but we are also working to ensure that the timing and content
of our speech production are aligned as seamlessly as possible with those of our interlo-
cutors. Utterances are formulated to fit into sequences of social interaction, and such
sequences are characterized by the orderly and finely timed transition of interlocutors
between speaker/hearer roles. This is the phenomenon of
For you to produce an irrelevant utterance, or one whose deployment is less than
impeccably timed, risks making an ‘unusual’ contribution, which in turn may cause
your interlocutor to infer messages you hadn’t intended (e.g. you’re getting impatient,
* We wish to thank the Max Planck Gesellschaft for supporting our work. The work of J. P. de Ruiter is
also supported by European Union grants IST-2001-32311 and FP6-IST2-003747. In addition, we wish to
thank Hanneke Ribberink for her assistance in running the experiments, and Anne Cutler, Nick Evans,
Stephen Levinson, Asifa Majid, and Tanya Stivers for their helpful comments on earlier versions of this
article. None are to blame for shortcomings of this study—we have not always followed their advice. Finally,
we wish to thank Brian Joseph, Norma Mendoza-Denton, and two anonymous referees for their help in
making this a better article.
LANGUAGE, VOLUME 82, NUMBER 3 (2006)516
you’re becoming hesitant). This constant ‘threat to face’ (Brown & Levinson 1978)
means that there is much at stake in getting the timing of speech production just right.
In their influential and widely cited treatment of the turn-taking problem, Sacks,
Schegloff, and Jefferson (1974) point out that to achieve the precise timing that enables
us to do no-gap-no-overlap turn transitions, we must be anticipating in advance the
precise moment at which a speaker’s utterance is going to come to a completion point.
This allows us to set the wheels of speech production in motion well before our ‘in
point’ arrives, and in turn (ideally) keep exactly one person talking at all times.
In a typical example of a brief phone exchange shown as a transcript in 1 (Rahman:
A:1:VM:(4), supplied by Tanya Stivers), transitions of turns at talk by the two speakers
involve virtually no overlap or silence. For every turn, the
FLOOR TRANSFER OFFSET
(FTO), defined as the difference (in seconds) between the time that turn starts and the
moment the previous turn ends, is indicated in square brackets; a positive value indicates
a gap (short silence) between the two successive turns, whereas a negative value indi-
cates the amount of overlapping speech.
(1) Mat: ‘lo Redcah five o’six one?,
Ver: [Ⳮ0.15] Hello Mahthew is yer mum the:hr love.
Mat: [Ⳮ0.13] Uh no she’s, gone (up) t’town,h
Ver: [Ⳮ0.24] Al:right uh will yih tell’er Antie Vera rahn:g then.
Mat: [ⳮ0.03] Yeh.
Ver: [Ⳮ0.13] Okay. She’s alright is she.
Mat: [Ⳮ0.10] Yeh,h
Ver: [Ⳮ0.07] Okay. Right. Bye bye luv,
Mat: [Ⳮ0.02] Tara, .h
The unyielding imperative of participants in a conversation is to minimize both the
amount of speech produced in overlap (i.e. avoid having two or more people speaking
at the same time) and the amount of silence between successive turns. Our working
assumption is that the one-speaker-at-a-time rule is operative in all informal conversa-
tional settings. It is important to understand that to propose such an imperative or ‘rule’
is not to propose that all talk actually proceeds one speaker at a time. There are constant
departures from the rule (overlaps, gaps, and so forth, as is made explicit in the Sacks
et al. 1974 model; cf. Schegloff 2000:47–48, n. 1), and these departures can be exploited
for functional effect (since they are indeed treated by interactants as departures). These
departures may in addition mark differences in personal and cultural style (cf. Tannen
1985). We often encounter informal objections to the Sacks et al. 1974 model, or its
ilk, on the basis that in such-and-such a culture or social setting, different standards
seem to apply (a common one is that ‘in Language/Culture X, people talk in overlap
all the time’). Whether such claims are true remains an important empirical question.
There are to date no systematic studies of informal conversation that provide counterex-
amples to the claim of a one-speaker-at-a-time ‘design feature’ for the regulation of
conversational turn-taking (Schegloff 2000:2). Sidnell (2001) conducted an empirical
investigation of putative ‘contrapuntal’ conversation in Antigua (reported by Reisman
1974) but found the data to be compatible with the one-at-a-time model. Thus, as in
many domains of linguistic analysis, first intuitions turn out not to be supported by
empirical data. We strongly encourage further empirical attempts to describe and ac-
Nonspeech sounds like laughter that occurred at the beginning or end of turns were also assumed to be
part of that turn.
PROJECTING THE END OF A SPEAKER’S TURN 517
count for the distribution of timing of conversational floor transfer in other languages,
cultures, and social settings.
The present study begins with the observation that in our sizable data set from Dutch
two-party telephone conversations the vast majority of floor transfers have very little
gap or overlap. Consider our analysis of the FTO values of all 1,521 speaker transitions
in a corpus of phone conversations collected in our laboratory (see below for details
on data collection). In Figure 1, the distribution of the FTO values are shown: 45% of
all speaker transitions have an FTO of between ⳮ250 and 250 milliseconds, and 85%
of them are between ⳮ750 and 750 milliseconds.
1. Floor transfer offset distribution of 1,521 speaker transitions.
People accomplish this smooth temporal alignment of their talk by anticipating or
‘projecting’ a moment in time at which transition of speaker/hearer roles will be possi-
ble, relevant, and appropriate, allowing them to gear up in advance to begin talking at
just the right moment. It is important to emphasize that listeners do not wait until they
the end of a speaker’s turn; rather, they
this moment. An alternative
account of turn-taking by Duncan (1974) and Duncan and Fiske (1977) accounts for
the regulation of speaker transition by assuming that explicit
In the case of turn transitions like the ones in example 1 above, a smooth speaker
change would involve a signal from the speaker, followed by a signal from the listener,
followed by yet another signal from the speaker (see Figure 1 in Duncan 1974:178).
But for the smooth speaker changes under investigation, people simply do not have
the time to exchange three signals to negotiate a transfer of floor (see also Cutler &
Pearson 1986). Rather, listeners in a conversation must rely on anticipating the moment
that the current speaker will complete a turn, making the ‘floor’ available to be taken
up by another speaker. Sacks and colleagues refer to this moment as a
LANGUAGE, VOLUME 82, NUMBER 3 (2006)518
(henceforth, TRP), coinciding with the end of a turn constructional
unit (i.e. the minimal unit from which turns at talk may be built) produced by a current
speaker. Transition relevance in the sense used by Sacks et al. 1974 is a complex notion,
where relevance of transition includes reference to the social action being performed
by the turn at talk in question (e.g. telling, complaining, requesting, answering). The
ability to accurately anticipate the moment of a TRP’s occurrence presupposes the
ability to project a
(TCP). This ability to project is essential
to Sacks and colleagues’ ‘locally organized’ model of turn organization. Only if poten-
tial next speakers are able to accurately predict the end of a current speaker’s turn can
they plan their own turn to be accurately aligned in time with the current speaker’s
turn to the degree attested empirically. This implies that listeners—as potential next
speakers—face the task not only of comprehending the speech they are listening to,
but also, in parallel, preparing their next turn and predicting the moment of occurrence
of a TCP, where they should be already beginning to speak. In addition, they will need
to keep track of issues in the action-sequence structure to assess relevance and/or
appropriateness of turn transition, so as to judge when a detected TCP is a TRP. We
do not address this here.
Sacks and colleagues do not attempt to solve the question of what is involved psycho-
logically in carrying off accurate end-of-turn projection: ‘How projection of unit-types
is accomplished, so as to allow such ‘no gap’ starts by next speakers, is an important
question on which linguists can make major contributions. Our characterization in the
rules, and in the subsequent discussion, leaves open the matter of how projection is
done’ (1974:703). We now turn to a discussion of the candidate explanations.
. Several hypotheses have been formulated as to the informa-
tion sources that listeners use in projecting turn-completion points. Sacks and colleagues
(1974) suggested that syntax provides the main projection cue. Other cues that have
been suggested in subsequent research include pragmatic information (Ford & Thomp-
son 1996:151), pauses (Maynard 1989), and prosody (Couper-Kuhlen & Selting 1996a
and see references below).
Conversation-analytic work on the role of prosody in projection has identified spe-
cific intonational contours that often correspond with turn endings. Local, Wells, and
Sebba (1985) inspected the turn-final utterance you know in a dialect of English called
London Jamaican. They showed that the interpretation of you know as a turn-yielding
cue depended on the prosodic characteristics of the utterance. One of the prosodic cues
that corresponded with a turn-yielding you know was a falling pitch. Similarly, Local,
Kelly, and Wells (1986) found that a rising pitch contour was associated with turn-
yielding in Tyneside, another dialect of English, while Wells and Peppe
similar findings for Ulster English, a dialect of English that is special because it has a
final rising pitch pattern for declaratives.
Especially valuable are those studies that present exhaustive analyses of entire collec-
tions of turn transitions in corpora of naturally occurring conversation. Ford and Thomp-
Our consideration of the possibilities is restricted to properties of the speech signal, since we— like
Sacks and colleagues (1974)—are working with data from a nonvisual mode of conversation, namely tele-
phone calls. Given our current rudimentary state of understanding of the turn-taking problem, it is important
to keep matters simple by minimizing the variables at play. Since conversation typically occurs in a face-
to-face situation, it will be necessary in due course to give proper consideration to the role that nonverbal
behaviors such as eye gaze and hand gestures play in the organization of turns at talk (Kendon 1967, Goodwin
1981, Bolden 2003, Hayashi 2005, Rossano 2005).
PROJECTING THE END OF A SPEAKER’S TURN 519
son (1996) analyzed 198 speaker changes in twenty minutes of excerpts from multiparty
face-to-face conversation in American English. All excerpts were coded for points
of syntactic completion, points of intonational completion, and points of pragmatic
completion (i.e. the point where a ‘conversational action’ is complete, p. 151). Ford
and Thompson found that syntactic and intonational completion points frequently coin-
cided, and noted that every pragmatic completion point is an intonational completion
point, but not necessarily the other way around. Points in the analyzed conversation
where all three types of completion point coincided were termed
. It is at these complex TRPs, the authors claim, that turn transitions
most frequently occur. In their data, nearly three quarters of all speaker changes occur
at a complex TRP, and about half of the complex TRPs were accompanied by speaker
Another study, aimed specifically at investigating the role of intonational cues in
projection, was done by Caspers (2003), who studied a number of Map Task interactions
(Anderson et al. 1991). Like Ford and Thompson, Caspers also found a frequent coinci-
dence of syntactic completion points and intonational completion points. The main
result from Caspers’s study was that intonational cues are not generally used as turn-
yielding signals, but rather that rising pitch followed by a high level boundary tone
(H*_% in the ‘Transcription of Dutch Intonation’ system called ToDI;
ven 2005) is used as a turn-
signal, to ‘bridge’ syntactic completion points and
pauses longer than 100 ms. This contour is characterized by a rise well before the end
of an utterance, and a level pitch after this rise until the end of the turn. The idea is
that the speaker in such a case is signaling, by using this pitch contour prior to a pause,
‘Despite this pause, I’m not finished with my turn’.
In the studies discussed so far, the intonational cues that have been investigated in
connection with the projection of turn endings occurred
the end of turns. This has
two implications. First, these cues may be occurring too late in the speech to allow the
the end of the turn. Even in the case of the H*_% contour
described by Caspers (2003), which starts before the end of the turn, the listener has
to wait until the end of a turn in order to detect that the pitch contour has a level
boundary tone, and not a low boundary tone (H*L%). Second, the observation that
certain intonational phenomena
with turn endings does not mean that they
are used by the listener as anticipatory cues for projection.
The study by Schegloff (1996) is not vulnerable to the first part of this critique, as
he identified a pitch cue that occurs before the end of the turn. According to Schegloff,
a high pitch peak can signal that completion of the turn will occur at the next syntactic
completion point. In this case, listeners would have ample time to anticipate the end
of the turn after perceiving the pitch cue. Auer has called this type of explanation the
‘filter model’ (1996:85). In the filter model, intonation is used by listeners as a filter
to decide which syntactic completion points are turn endings, and which are not. How-
ever, Auer argues on the basis of data from conversations in German that syntax and
essential for listeners to determine whether a turn is completed or
A few experimental studies have also addressed this issue. Schaffer (1983) used
excerpts from conversations and applied low-pass filtering to create versions of these
fragments in which the speech was unintelligible, but the intonation was preserved.
To improve readability, we have used the underscore character to denote what is a (meaningful) space
in a ToDI string.
LANGUAGE, VOLUME 82, NUMBER 3 (2006)520
She then asked subjects to listen to the fragments and indicate whether they thought
the fragments were turn-final or not. Schaffer found that lexicosyntactic information
was a much more consistent cue than intonation for end-of-turn detection; subjects
listening to unfiltered stimuli showed more (significant) agreement than subjects listen-
ing to stimuli from which the lexicosyntactic information, but not the intonational
contour, was removed.
Cutler and Pearson (1986) created dialogue fragments by having speakers read written
dialogue scripts, and used the recordings of those readings in an experiment in which
subjects had to indicate whether the fragments they heard were turn-final or turn-medial.
They found some evidence for rising or falling intonational contours signaling turn-
finality, but also noted that ‘many of the utterances that our listeners found ambiguous
also had upstepped or downstepped pitch’ (Cutler & Pearson 1986:152). Beattie, Cutler,
and Pearson (1982) performed a similar experiment on fragments from interviews with
Margaret Thatcher, to find out why she was apparently interrupted so often. They
concluded that Mrs. Thatcher’s use of sharply dropping intonational contours at posi-
tions that were not turn-final often misled listeners into thinking that she was finished
with her turn.
To summarize, some studies have argued that syntax is the main information source
for accurate end-of-turn projection, while others have suggested that intonational con-
tour also provides essential cues. A complication in evaluating these latter claims is that
they rely on an observed
of intonational patterns and speaker changes. In
natural conversational data, syntactic and intonational structure are correlated (Ford &
Thompson 1996, Caspers 2003), which makes it difficult to estimate their relative
importance as cues for projection. To reiterate a point already made above, correlation
does not by itself imply causation, so establishing a correlation between intonational
cues and speaker change does not necessarily mean that the intonational contours are
by listeners. Knowing the location of a TCP in its conversational context
does not inform us whether it is intonational contours, syntactic cues, pragmatic cues,
pauses, or some combination of these that listeners have used in anticipating the TCP.
Other researchers (Beattie et al. 1982, Schaffer 1983, Cutler & Pearson 1986) have
tried to approach this problem experimentally by having subjects make judgments on
turn-taking issues. But these judgments were off-line, meaning that subjects were not
under time pressure to make their judgment. In reality, listeners have to anticipate TCPs
in real time, with only fractions of a second to process the speech signal.
HE EXPERIMENTAL TASK
. We have endeavored to better approximate the condi-
tions under which projection is actually done by having listeners do real-time projection
of turn endings, using real conversational data as stimulus materials. We presented
fragments of such conversations to subjects and asked them to press a button at the
moment they thought that the speaker they were listening to had finished his or her
turn. We explicitly instructed our subjects to
this moment, because we did
not want to have the equivalent of a reaction-time experiment where subjects
to a detected event. This on-line method has the advantage that it mimics the temporal
constraints in which projection takes place in nature. Subjects heard the fragments only
once and had to anticipate TCPs on the fly.
To be able to tease apart the various candidate cues for projection from among those
features that previous descriptive studies have shown tend to cluster at TCPs, our
experimental set-up involved manipulation of the speech signal. We needed a way of
making a controlled assessment of the relative contribution of three possible cues for
PROJECTING THE END OF A SPEAKER’S TURN 521
projection: (a) lexicosyntactic information, (b) intonation, or pitch, and (c) amplitude
(volume) envelope of the signal. To this end we created stimuli for the on-line projection
task by selecting turns at talk from recorded conversations and creating different ver-
sions of each, varying the information content in the signal: presence versus absence
of lexicosyntactic information; presence versus absence of pitch information; presence
versus absence of rhythmical or amplitude-envelope information. The logic is simple.
If a certain source of information is used in projection, the absence of that information
should have a detrimental effect on the accuracy and/or consistency of the prediction
performance of the subjects.
. For maximum ecological validity, our stimulus materials
were taken directly from recordings of natural conversation. We chose not to construct
stimuli using actors, since this may result in omission of natural cues or addition of
unnatural ones (see also Cutler & Pearson 1986). However, we had to devise a way
not only to get the highest quality audio recordings, but also to avoid cross-talk between
the recordings of the two speakers. To accomplish this, we recorded the conversational
data using two soundproof cabins. Each cabin had a microphone on a desk in front of
the subject and a set of closed headphones. A digital tape recorder was used to record
the speech of the subject in cabin A on the left track of a stereo audio recording, and
the speech of the subject in cabin B on the right track of the recording. The two speakers
could each hear both themselves and their interlocutor in their headphones, giving the
effect of a normal phone conversation. The recordings were of top acoustic quality,
and there were no traces of the speech of one subject on the recording of the other.
. Sixteen native speakers of Dutch participated in the stimulus collec-
tion, with eight dyadic interactions (one male-male, three female-female, four female-
male). The speakers in each pair were friends. For each dyad, once the two were seated
in their respective soundproof cabins, we explained that they could talk to each other
via the headphone and microphone, and told them that we were testing our new sound
equipment and just needed some random speech samples, so they could freely talk to
one another about anything they liked. Each dyad was then recorded for fifteen minutes,
resulting in a total of two hours of recorded conversation. The recorded conversations
were lively and appeared in all aspects like informal phone conversations between two
friends. Thus, there is nothing about the content or composition of these interactions
that we would not expect in a natural setting.
. All recorded conversations were transcribed to a high level
of detail, including overlapping speech, turn-beginnings and endings, in-breaths, laugh-
ter, pauses, and back-channel signals (Yngve 1970) or ‘continuers’ (Schegloff 1982)
such as m-hm.
We then selected a set of turns to use as stimuli in the TCP projection experiments.
These turns all contained at least five words. This was to ensure that the subjects in
the experiments would have some reasonable amount of signal to base their response
on. Short turns contain only limited amounts of lexical and pitch information which
could lead to the undesired possibility that we would measure only how fast the subjects
were aware that the stimulus had already finished.
We selected 108 target turns and an additional 54 turns for practice purposes. The
set of target turns was a random selection from our corpus, although we made sure
that the target turns displayed sufficient range with respect to duration and amount of
overlap with the turn that followed it in the original conversation (roughly, one third
LANGUAGE, VOLUME 82, NUMBER 3 (2006)522
had some overlap in the original conversation, one third had some silence between the
turns (i.e. ‘negative overlap’), and one third were almost perfectly aligned). We also
made sure that there was a balance between sex of the speaker (50% male and 50%
female). The total number of different speakers in the selected turns was fourteen (eight
female and six male speakers).
Table 1 presents some descriptive statistics of the target turns. For FTO (see also
the discussion of Fig. 1 above), a negative value indicates the duration (in milliseconds)
of the overlapping speech with the following turn (which is not audible in the fragment
that is presented to subjects in the experiment), whereas a positive value indicates the
duration of the silent period between the target turn and the following turn.
MINIMUM MAXIMUM MEAN
Duration (ms) 929 9952 2904 1899
FTO (ms) ⳮ3740 1360 ⳮ78 798
Number of words 5 39 13.1 7.85
Mean pitch in Hz 99 300 163 49
(semitones ref 100 Hz) (0.0) (19.0) (8.0) (4.5)
Std. dev. of pitch within 0.03 5.77 2.49 1.14
Pitch range within turn 0.04 28.80 11.70 5.15
1. Descriptive statistics of the selected target turns.
These turns were extracted into individual sound files using the phonetic-analysis pro-
gram Praat 4.2 (Boersma & Weenink 2004). We extracted only the turn itself, and not
the sound from the other channel in which the speech of the interlocutor’s subsequent
turn had been recorded. Thus, only one speaker would be heard on the resulting stimulus
sound file, even if the other speaker was speaking (i.e. in overlap) in the original
REPROCESSING OF EXPERIMENTAL STIMULI
. Five versions were created of every
turn fragment. A NATURAL version was the original fragment as is. A NO-PITCH
version was created by ‘flattening’ the pitch (F0) contour by PSOLA resynthesis using
Praat 4.2: the pitch was set to the mean pitch value of the original fragment, such that
the pitch contour was completely horizontal. A NO-WORDS version was created by
low-pass filtering the original fragment at 500 Hz (50 Hz Hanning window). With the
NO-WORDS fragments, it is impossible to identify the words, but the original pitch
contour remains clearly distinguishable. In a NO-PITCH-NO-WORDS version we ob-
tained fragments with a flat pitch contour
unidentifiable words by applying both
the low-pass filtering and the PSOLA resynthesis. (We applied the low-pass filtering
after the PSOLA procedure, because the latter procedure uses information in the higher
frequencies for the pitch estimation.) A NO-PITCH-NO-WORDS-NO-RHYTHM ver-
sion, which we call the NOISE version for brevity, was created by generating a sample
of constant noise with the same duration and frequency spectrum as the original frag-
ment. This was achieved by convolving the speech stimulus with white noise. This
version was used as a comparative baseline, to see whether the amplitude-envelope
information that is still present in the NO-PITCH-NO-WORDS stimuli was an effective
cue for projection. All five versions of each fragment were equated in loudness using
the sone scale.
The low-pass filtered stimuli generally tend to sound softer because their acoustic energy (which is
concentrated in a few critical bands) suffers more from lateral and forward masking than the natural and
PROJECTING THE END OF A SPEAKER’S TURN 523
Below, we report on two experiments. In the first experiment, subjects were presented
with the NATURAL, NO-WORDS, and NO-PITCH stimuli, and in the second, a new
group of subjects was presented with the NO-WORDS (this condition was replicated),
NO-PITCH-NO-WORDS, and NOISE stimuli. The reason we ran two experiments
instead of one experiment containing all five conditions was that it would have compli-
cated the design considerably, leading to a much smaller number of trials in every
stimulus/condition design cell. Instead we wanted to collect enough data to derive
individual response distributions for every stimulus in every condition.
NATURAL, NO-PITCH, AND NO-WORDS FRAGMENTS
. Every subject was presented with three blocks of thirty-six stimuli each,
taken from one of the three manipulation groups NATURAL, NO-WORDS, and NO-
PITCH. There were six experimental lists; three were permutations of the order in which
the blocks were presented, respectively NATURAL (NAT)—NO-PITCH (NP)—NO-
WORDS (NW), NW—NAT— NP, and NP—NW—NAT. Within each block, six prac-
tice trials were followed by thirty-six randomly selected target stimuli. These were
selected such that all 108 target stimuli would be presented in one of the three presented
experimental blocks, and none of the 108 stimuli would be presented twice within the
same experiment. The remaining three lists were the same as the first three lists, only
with the sequential presentation order of the stimuli reversed. These six lists were
created in order to counterbalance both potential effects of block order and the order
of the stimuli within blocks and over the entire experiment. The order of the practice
trials was also reversed for these three lists (but they were, of course, placed before
their respective block in the experimental presentation sequence).
. Sixty native speakers of Dutch (thirty-nine women and twenty-one
men) participated in the experiment. They were assigned randomly to one of the six
experimental lists (ten subjects per list). Subjects were seated in a soundproof cabin
and given written instruction to listen to the fragments that would be presented to them
through closed headphones, and press a button in front of them at the moment they
thought the speaker would be finished speaking (Dutch: is uitgesproken). The instruc-
tion encouraged the subjects to try to
this moment, and not wait until the
fragment stopped playing. They were informed that there were three blocks of trials,
and that in some or all of the blocks, the fragments were manipulated acoustically.
Subjects were then presented with the six practice trials of the first block. After the
practice trials, the first block of stimuli was presented, and after that the same sequence
(first practice, then experimental trials) occurred for the second and third block. Every
experimental trial consisted of a visual ‘countdown’ from 3 to 1 presented on the
computer screen in front of them, followed by the acoustic presentation of the experi-
mental fragment. An important aspect of the procedure was that as soon as a subject
pressed the button, the sound would immediately cut out. This was necessary to avoid
giving subjects feedback about their performance. If the subjects still heard the fragment
playing after having pressed the button, they might become more conservative, and
develop a strategy of waiting for the end of the fragment before pressing the button.
Button-press times relative to stimulus onset were recorded by computer. If a subject
had not pressed the button by 2,000 ms after stimulus offset, a time-out was recorded.
pitch-flattened stimuli. Therefore the sone scale was used instead of the acoustic energy, to compensate for
differences in perceived loudness.
LANGUAGE, VOLUME 82, NUMBER 3 (2006)524
NO-WORDS, NO-PITCH-NO-WORDS, AND NOISE FRAGMENTS
. This experi-
ment, with sixty new subjects (forty-four women and sixteen men), differed from experi-
ment 1 only in having the three conditions NO-WORDS (replicated from experiment
1), NO-PITCH-NO-WORDS, and NOISE.
ESULTS AND DISCUSSION
. We wanted to know whether the experimental para-
digm was tapping into the cognitive process of anticipating TCPs. The following find-
ings demonstrate that indeed it was.
First, for the NATURAL (unmodified) stimuli, response times were very accurate.
Subjects were highly skilled at anticipating the moment at which speakers would finish
their turn at talk. The average BIAS (defined as response time recorded from stimulus
onset minus duration of target stimulus) within the NATURAL condition was ⳮ186
ms, indicating that on average, subjects pressed the button 186 ms before the end of
the target stimulus. Given that the average duration of the stimuli was 2,904 ms (and
the average BIAS only 6% of that), this is remarkably accurate, and reflects what is
found in natural conversation. In Figure 2 below, the BIAS of all trials in the NATURAL
condition are plotted as a distribution, in the same format as Fig. 1. Similarly to the
FTO values in Fig. 1, a negative BIAS value indicates a response before the fragment
was finished, and a positive BIAS value a response that occurred after the fragment
was finished. A comparison of Fig. 2 with the natural data in Fig. 1 reveals that the
projection performance of our experimental subjects had the same mode (at 0) and the
same distributional shape as the natural data. Of course, in contrast to the participants
in the natural conversations, the subjects in our experiment did not need to prepare and
execute a verbal response to the fragment they were listening to, which may explain
2. BIAS distribution of responses in NATURAL condition.
PROJECTING THE END OF A SPEAKER’S TURN 525
the higher accuracy (reflected in a higher proportion of values in the FTO ⳱0 bin)
of these button-press projections when compared to the projection performance of the
subjects in the real conversations.
Second, the average BIAS was negative in all conditions (BIAS data for the other
conditions are discussed below). This indicated that subjects did not, generally, wait
for the end of the signal before responding. In some of the trials, the subject responded
after the turn was complete, but the negative average BIAS shows that this was not a
general strategy. In other words, the subjects were able to
Third, the distribution of responses for many of the longer target turns indicated the
existence of multiple TCPs in the stimuli. Given that we had many subject responses
to each individual target turn (forty per stimulus in the NO-WORDS condition, and
twenty in the other conditions), embedded TCPs that were present in the target stimulus
before offset should result in peaks in the response distribution at the earlier TCP as
well as the TCP at the end of the stimulus. This is exactly what happened with the
majority of the longer stimuli. As a representative example, see Figure 3, in which the
waveform of the target turn, the word segmentation, the pitch contour, and the response
distributions are displayed on one horizontal time axis. The top panel shows the response
distribution for NATURAL (solid line) and NO-PITCH (dotted line) conditions, while
the second-to-top panel shows the response distribution for NO-WORDS (solid line)
and NO-PITCH-NO-WORDS (dotted line) conditions.
3. Example stimulus with pitch contour and response distributions. From bottom to top, panels
represent: A: transcription with each word directly under the piece of speech it occupies, B: sound
wave, C: measured F0 (pitch) contour, D: response distribution of the NO-WORDS (solid)
and NO-WORDS-NO-PITCH (dotted) conditions, and E: response distribution of
NATURAL (solid) and NO-PITCH (dotted) conditions.
LANGUAGE, VOLUME 82, NUMBER 3 (2006)526
The transcription and English gloss of the fragment displayed in Fig. 3 is given in 2.
(2) Neeja ‘kheb’m officieel isie no(g) van mij maar eh .hh⳱
No yes I have it officially is it still mine but eh .hh⳱
⳱‘kheb er al eh vanaf ‘t voorjaar⳱
⳱I have there already eh from the spring⳱
⳱nie meer naar omgekeken⳱
⳱not anymore cared for⳱
⳱eigenlijk vanaf de winter al niet meer.
⳱actually from the winter on not anymore.
Figure 3 reveals that the response distribution for the NATURAL condition is bimodal,
with a peak not only at the end, but also around 2,000 ms before offset, which corre-
sponds with the third syntactic completion point in the turn.
(The response distributions
of the other two conditions are discussed below.)
CCURACY OF SUBJECTS’ RESPONSE: BIAS
. We now turn our attention to the
recorded response BIAS in the different conditions. Figure 4 presents an overview of
the average BIAS per condition. In this and all subsequent analyses,
valid responses are included, including early responses that may have been caused by
embedded TCPs as mentioned above. Due to the nature of the task, longer fragments
generally result in a more negative bias than shorter fragments. This is a simple stochas-
tic effect: with the longer fragments the subjects have a longer period during which
they could press the button too early than with the shorter fragments. This results in
a negative correlation between BIAS and DURATION (r⳱ⳮ0.636, p⬍0.001).
Because of this relationship between DURATION and BIAS, all of the statistical tests in
this study are performed
. This means that in comparing the different
4. Average BIAS of responses per condition.
* indicates statistical significance at the 0.05 level.
A free translation of this fragment: ‘No, yes, I have . . . officially it is still mine, but uh actually I haven’t
given it any attention anymore since spring, in fact since winter.’
We defined syntactic completion points in the same way as Ford and Thompson (1996:143–45).
Interestingly, this relationship is also present in our conversational data. The longer the (first) turn, the
more negative the FTO. In our conversational data, the correlation between DURATION and FTO is ⳮ0.136
PROJECTING THE END OF A SPEAKER’S TURN 527
conditions, every condition always contains the same 108 stimuli. This way, we ex-
cluded the possibility that the relationship between DURATION and BIAS influenced
An ANOVA shows a significant main effect for presentation condition (by subjects:
F1(4,353) ⳱73.37, MSE ⳱165,610, p⬍0.001; by items: F2(4,428) ⳱64.95, MSE
⳱330,938, p⬍0.001). A post-hoc analysis (Tukey HSD, ␣⳱0.05) indicated that
all differences between individual conditions were significant, with the key exception
being the difference between the NATURAL and the NO-PITCH condition, which was
not significant in either item- or subject-analysis. A general linear model analysis re-
vealed no significant effects of the independent variables SPEAKER-SEX or SUB-
JECT-SEX on BIAS, nor any significant interactions of SPEAKER-SEX or SUBJECT-
SEX with CONDITION. The partial correlation between FTO and BIAS, controlling
for DURATION, is not significant in any of the conditions.
The most important finding is that removing pitch information from the stimuli had
no influence on projection accuracy. By contrast, removing lexical content by the low-
pass filtering procedure in the NO-WORDS condition
have a strong detrimental
effect on projection accuracy.
The significant difference in BIAS between NO-WORDS and NO-PITCH-NO-
WORDS suggests that when no lexicosyntactic information is available, pitch can
sometimes be used as a ‘turn-keeping’ cue in the presence of pauses in the signal.
Grosjean and Hirt (1996) report a similar finding in experiments where subjects had
to predict the end of sentences in a
task (Grosjean 1980), noting that ‘It is only
when higher-level information is no longer present...that prosodic information is
made available or is called into play’ (Grosjean & Hirt 1996:129). In the NO-PITCH-
NO-WORDS condition, the main acoustic cue available for projection in this experi-
ment is the amplitude envelope. A silence of a certain duration will, in the absence of
any lexicosyntactic information, often be interpreted as the end of the turn. If a turn-
keeping pitch contour, for instance the H* %contour as described by Caspers (2003),
is present, this may delay the response. Inspection of the response distributions in the
NO-PITCH-NO-WORDS condition confirmed that pauses in the speech signal were
indeed the main determinant for the response distributions in the NO-PITCH-NO-
WORDS condition. In the data from Fig. 3, the response distributions of the conditions
without lexicosyntactic information (NO-WORDS and NO-PITCH-NO-WORDS; see
the arrow in the D panel) both peaked exactly at the pause in the speech. In Figure
5, another illustrative stimulus is depicted together with the corresponding response
distributions. The transcription and English gloss of the fragment is given in 3.
(3) Ah en waren er nog eh (0.5 s) bijzondere mensen die eraan ⳱
Ah and were there yet eh (0.5 s) special people who there in ⳱
⳱meededen of nie.
⳱participated or not.
In the NO-PITCH-NO-WORDS condition, the majority of the responses occur
around the 500 ms pause (at about 2.3 s). In contrast, in the NO-WORDS condition,
where the pitch contour is still present, the majority of responses occur at the end of
the fragment. This indicates that having access to the pitch contour prevented the
subjects in the NO-WORDS condition from responding to the pause, whereas the sub-
Partialing out the effect of DURATION is necessary in this analysis, because FTO is negatively correlated
with DURATION (r⳱ⳮ0.207, p⬍0.05) for the fragments used in the experiment.
Free translation: ‘ah, and were there uh any special people who participated in it or not?’
LANGUAGE, VOLUME 82, NUMBER 3 (2006)528
5. Example stimulus with pause. From bottom to top, panels represent: A: transcription with each
word directly under the piece of speech it occupies, B: sound wave, C: measured F0 (pitch) contour,
D: response distribution of NO-WORDS (solid) and NO-WORDS-NO-PITCH (dotted)
conditions, and E: response distribution of NATURAL (solid) and NO-PITCH
jects in the NO-PITCH-NO-WORDS condition were ‘fooled’ by the pause. The high/
level pitch contour serves as a turn-keeping cue to ‘override’ pauses in the signal (cf.
Lexical information can also ‘override’ pauses in the speech. In the fragments in
both Figs. 3 and 5, the pauses are preceded by the word eh. Clark and Fox Tree (2002)
investigated the function of uh (the English equivalent of the Dutch eh) in conversation.
Although their analysis did not address speaker changes in conversational turn-taking,
Clark and Fox Tree raise the possibility that English uh could be used as a turn-keeping
cue, that is, as signaling ‘I’m having trouble producing this utterance but I’m not finished
yet, so wait’ (see also Jefferson 1973). Our investigation provides strong support for
a turn-keeping function of eh in conversation. In our stimulus set, there are fifteen
fragments containing pauses longer than 200 ms, and ten of them (two thirds) are
immediately preceded by eh. Furthermore, in the NATURAL and NO-PITCH condi-
tions (where the eh could be heard and identified), the likelihood of a response peak
at a pause without a preceding eh was twice as high as the likelihood of a peak in the
response distribution at a pause that
preceded by eh. In other words, in the condi-
tions where eh was identifiable, its presence prevented the pause from being interpreted
as a TCP.
The finding that the removal of pitch information does not have any effect on projec-
tion accuracy seems, at first sight, to be at odds with the finding by Beattie and col-
leagues (1982), who investigated why Margaret Thatcher was interrupted so often
PROJECTING THE END OF A SPEAKER’S TURN 529
during interviews. They found that a rapidly falling pitch in her voice led her interview-
ers to ‘mis-project’ and take up the next turn, although apparently Mrs. Thatcher had
not yet finished. This shows that under certain circumstances, it
possible for pitch
contours to suggest the presence of a TCP (see also Schegloff 1996). We inspected the
individual response distributions for all of our experimental stimuli and located seven
fragments for which the mode of the response distribution for the NATURAL data
precedes the mode of the NO-PITCH distribution by 250 ms or more. So for these
cases, when the stimulus contained pitch information, subjects responded at least 250
than when the stimulus did
contain pitch information. In all of these
seven cases, there was either a noticeable pitch rise (two cases, mean pitch velocity
Ⳮ22 semitones/s) or fall (five cases, mean pitch velocity ⳮ12 semitones/s), and these
pitch contours occurred at syntactic completion points. Both in the Beattie et al. 1982
and our own studies, a rapidly changing pitch could mislead listeners into projecting
the end of the turn at an early syntactic completion point. In our study, the presence
of these cues led to early responses to these fragments in the NATURAL condition.
Intriguingly, these fragments did not result in early responses in the NO-WORDS
condition, where the subjects had full access to the pitch contour. If rapidly falling or
rising pitch contours
indicated the presence of TCPs, they should have
led to early responses in the NO-WORDS condition. The fact that they did not means
that these pitch cues indicate the presence of a TCP only if lexicosyntactic information
is also available (as was the case in all of the conditions in the study by Beattie and
colleagues). This suggests that lexicosyntactic information is, again, the crucial source
of information, even in the case of ‘mis-projections’.
A final BIAS effect to discuss is that even in the NO-PITCH-NO-WORDS condition,
projection accuracy was still better than chance, chance performance being represented
by the NOISE baseline in which the subjects had absolutely no information for project-
ing. This indicates that the rhythmical properties of the fragments in the NO-PITCH-
NO-WORDS stimuli still contained some cues for projection.
Our main finding from the analysis of BIAS is that removing the pitch contour from
the original stimulus did not have an effect on subjects’ projection performance. By
contrast, removing the lexicosyntactic content (while retaining the pitch contour) did
significantly deteriorate projection performance. In those cases where the lexicosyntac-
tic information was removed from the speech, intonation was of some help in signaling
that longer pauses in the speech were not turn endings.
ONSISTENCY OF SUBJECTS’ RESPONSES: ENTROPY
. Another way to investigate the
relative role of pitch and lexical information in projection, in addition to the analysis
of the BIAS presented above, is to inspect the
with which subjects re-
sponded to the stimuli in different conditions. That is, we do not ask how
they were, but rather how much
there was among subjects with respect to
when they responded. The subjects may have been ‘wrong’ in their projection compared
to the original interlocutor, but the higher the agreement among them, the more likely
it is that they have used the cues present in that particular condition. A straightforward
way to estimate this level of agreement from the data is to compute the standard devia-
tion of all responses for every stimulus/condition pair. This, however, has a serious
disadvantage. As has been shown in the analysis above, there can be multiple TCPs
Reanalysis of BIAS excluding the seven stimuli that contained these pitch cues led to an identical pattern
in the overall results.
LANGUAGE, VOLUME 82, NUMBER 3 (2006)530
in one turn fragment, leading to multimodal response distributions. The further apart
these TCPs are in time, the higher will be the standard deviation of the responses, even
though the agreement with respect to the perceived location of those multiple TCPs
might have been high. To circumvent this problem, we computed for every stimulus/
condition pair the
as defined by Shannon (1948),
a measure of uncertainty
that does not have the disadvantage mentioned above. The entropy measure is insensi-
tive to the actual temporal distance between these multiple modes. However, it is
sensitive to the level of agreement. If all responses were to occur within the same
interval, the entropy would be zero. The more the responses are distributed over different
intervals, the higher the entropy is. Therefore, the entropy gives us a reliable estimate
of the amount of agreement among subjects, without being biased by the temporal
distance between possible multiple TCPs in the fragments.
6. Shannon entropy of responses per condition.
* indicates statistical significance at the 0.05 level.
In Figure 6, the average Shannon entropy (using a bin-width of 250 ms) is shown
for every condition. Again, there is a main effect of condition (F2(4,428) ⳱43.79,
MSE ⳱0.206, p⬍0.001).
A post-hoc analysis (Tukey HSD, ␣⳱0.05) indicated
that all differences between conditions were significant with two key exceptions: the
difference between NATURAL and NO-PITCH and the difference between NO-
WORDS and NO-PITCH-NO-WORDS. These differences were not significant.
This result supports the above interpretation of the BIAS data that the removal of pitch
information from the turn fragments does not have a detrimental effect on projection
performance: the consistency of subjects’ responses was unaffected by the removal of
pitch information, both for the conditions with and without lexicosyntactic information.
A similar conclusion was reached by Schaffer (1983:253), who concluded that ‘Syntac-
tic and lexical characteristics appear to be used much more consistently as cues to turn
status’. However, Schaffer did not have an experimental condition equivalent to our NO-
PITCH condition, where the lexicosyntactic information is present but the intonational
contour is not. Therefore, subjects listening to the unfiltered fragments in Schaffer’s
experiment always had
information available than those listening to the filtered
), where Nis the number of bins, and p
is the proportion of samples in
Because entropy distributions can be computed for every stimulus only over entire response distributions,
only a by-item statistical analysis could be performed here.
PROJECTING THE END OF A SPEAKER’S TURN 531
fragments, and this could be an alternative explanation of her finding. Our study ruled
out this alternative explanation by varying the presence of intonation and lexicosyntactic
The entropy results also provide support for the above explanation of the difference
in BIAS between the NO-WORDS and the NO-PITCH-NO-WORDS conditions. If
subjects did perceive the pitch contours present in the NO-WORDS fragments to be
turn-keeping cues when there was a pause in the fragment, this should have affected
the BIAS (as it did) but not the agreement about where the TCP was. Subjects in
the NO-PITCH-NO-WORDS condition may have been less accurate (hence the larger
absolute BIAS), but their level of agreement was nevertheless the same, because they
were all ‘fooled’ in the same way by the same pauses.
. In our experimental simulation of the conditions of
end-of-turn projection in Dutch conversation, subjects demonstrated the same remarka-
ble ability as speakers in natural conversation to accurately anticipate the moment of
end-of-turn completion. Different experimental conditions revealed that in order to
achieve this accuracy of projection, what they needed was the lexicosyntactic structure
of the target utterances. When we removed pitch contour from the target utterance,
leaving the lexicosyntax intact, this had
on hearers’ ability to accurately
project turn endings. And when we removed the lexicosyntactic content, leaving the
original pitch shape in place, we saw a dramatic decline in projection performance.
The conclusion is clear: lexicosyntactic structure is necessary (and possibly sufficient)
for accurate end-of-turn projection, while intonational structure, perhaps surprisingly,
is neither necessary nor sufficient.
This conclusion raises a host of further questions for subsequent research. For in-
stance, what will account for the observed differences in function of intonation and
lexicosyntax? Lexicosyntactic information is symbolic, conventionally coded, and hier-
archically structured. By contrast, features of the speech signal such as intonational
contours (in Dutch at least), pauses, and effects of their placement are iconic and
indexical, and therefore more likely to be language-independent (see e.g. Ohala 1983
for pitch). There is thus a strong asymmetry in the informational roles of lexicosyntax
and intonation: symbolic information strongly determines the interpretation of associ-
ated nonsymbolic information.
And while the functions of intonation display rich
variety, this variety is not infinite. Our study is a case in point: intonation does not give
Dutch listeners what they need in order to predict when a current speaker’s utterance is
going to end. With its high degree of formal flexibility, signaling a wide variety of
expressive functions, it may be that for the very task of end-of-turn projection, intonation
just cannot compete with the greater restrictiveness, and therefore greater predic-
tiveness, of syntax. While many people seem attached to the idea that it should play
a role in turn organization, we think it is only to be expected that there will be limits
to the range and number of functions of intonation.
A second line for further research concerns the precise mechanism by which lexico-
syntactic structure signals the timing of turn completion. What is it about Dutch lexico-
syntax that enables listeners to project turn completion so accurately in free conversation
(see Fig. 1, above) as well as in our experimental replication of the same task (Fig. 2)?
One avenue is to investigate languages whose grammatical structures differ dramatically
This is amply illustrated in research on co-speech hand gestures, which are often impossible to interpret
without the speech that accompanies them, but effortless to interpret when they are accompanied by speech
(McNeill 1992, Kendon 2004).
LANGUAGE, VOLUME 82, NUMBER 3 (2006)532
from Dutch, especially where conventional syntactic structures provide less constraining
information, for example where there is different and/or more variable word order, or
widespread ellipsis. It may turn out that lexicosyntactic structure is universally a better
source for end-of-turn projection. To find out why this might be so, one needs to
carefully consider the informational properties of unfolding syntactic sequences and
the complex relationship between the trajectory of information supply through the
course of a syntactic sequence and the available degrees of expressive freedom at any
given point. There are conceptual payoffs from this line of inquiry. Just posing the
question of how the emerging yet unfinished structure of an utterance can allow listeners
to project its completion point forces us to think of lexicosyntax as a temporally unfold-
ing structure which displays (often ambiguous) information about itself through the
course of its development. This truly temporal view of lexicosyntax is alien to most
corners of descriptive/typological linguistics,
but it is de rigueur in the two rather
disparate empirical approaches to language that we bring together in this study: experi-
mental psycholinguistics (e.g. Vosse & Kempen 2000) and conversation analysis (e.g.
Goodwin 1979, 2006).
The turn-taking problem forces us to think about syntax and cognition in new ways.
Grammar is not just a means for semantic representation of predicate-argument relations
(or different construals thereof; Langacker 1987, Wierzbicka 1988, Lambrecht 1994,
Goldberg 1995). Syntactic structure is also an inherently temporal resource for listeners
to chart out the course of a speaker’s expression, and to plan their own speech accord-
ingly. In anticipation of this, a speaker may exploit available lexicosyntactic options
to actively foreshadow structure, thereby manipulating the unfolding course of the
interaction itself, and literally controlling the interlocutor’s processes of cognition and
action. Thus, the virtues of studying conversation are not just that we are seeing language
in its natural home (Thompson & Hopper 2001:27), but that we are seeing how it may
strategic deployment in social interaction (Sacks et al.
1974). These considerations add to the range of factors that may help us understand
why languages are structured in just the ways they are.
In closing, we turn to methodological implications. The mode of research that has
made significant headway in the domain of conversation is the observational analysis
of spontaneously occurring conversation, the tradition known as conversation analysis
(Sacks et al. 1974, Levinson 1983, Heritage 1984). A key methodological insight is
that much evidence for underlying structure can be found in the observable details
of natural conversation itself. Each successive contribution reveals participants’ own
analysis of the conversational moves being made. But for this study, we needed a
different source of evidence. In natural conversation, each occasion of turn transition
is unique, supplying one occasion of response, and thus one and only one ‘indigenous
analysis’ of any given utterance. What we needed was an entire distribution of such
analyses by multiple speakers in response to a single stimulus. In isolating, varying,
and testing competing hypotheses, we endeavored to meet the challenge of balancing
experimental control and ecological validity. With stimuli culled from real interactions,
and with a real-time projection task, we simulated the psychological conditions of
Notable exceptions include ‘emergentist’ views such as Hopper 1998 and Ford et al. 2003. However,
since these approaches explicitly reject the idea that syntax is abstractly represented, they offer no account
for a listener’s ability to recognize and project larger oncoming structures on the basis of initial segments
of a currently emerging utterance. Nor, for the same reason, can they accommodate the idea that a speaker
may strategically deploy a given structure, designed so as to be recognized to be following a certain course.
PROJECTING THE END OF A SPEAKER’S TURN 533
projection in the wild. The results show that our measure was valid. Just as direct
inspection of conversational data will answer questions that experimental methods can-
not, our design yielded answers that could not be supplied by direct inspection of
conversational materials. The study demonstrates that conversation as an organized
structural and psychological domain is amenable to experimental investigation. We see
this kind of work not as an alternative to conversation-analytic or descriptive linguistic
work, but as a necessary counterpart, where the insights flow in both directions. May
the result signal a new trend in cross-disciplinary research on a topic at the very heart
of language: the structure of conversation.
; and R
The HCRC Map Task corpus. Language and Speech 34.351–66.
. 1996. On the prosody and syntax of turn-continuations. In Couper-Kuhlen &
Selting 1996b, 57–100.
; and M
. 1982. Why is Mrs. Thatcher inter-
rupted so often? Nature 300.744–47.
, and D
. 2004. Praat: Doing phonetics by computer (Version
4.2). Amsterdam: University of Amsterdam. Online: http://www.fon.hum.uva.nl/praat/.
B. 2003. Multiple modalities in collaborative turn sequences. Gesture
, and S
. 1978. Politeness: Some universals in lan-
guage usage. Cambridge: Cambridge University Press.
. 2003. Local speech melody as a limiting factor in the turn-taking
system in Dutch. Journal of Phonetics 31.251–76.
H., and J
. 2002. Using uh and um in spontaneous speaking.
, and M
. 1996a. Towards an interactional per-
spective on prosody and a prosodic perspective on interaction. In Couper-Kuhlen &
Selting 1996b, 11–56.
, and M
(eds.) 1996b. Prosody in conversa-
tion. Cambridge: Cambridge University Press.
, and M
. 1986. On the analysis of prosodic turn-taking cues.
Intonation and discourse, ed. by Catherine Johns-Lewis, 139–55. London: Croom
. 1974. On the structure of speaker-auditor interaction during speaking
turns. Language in Society 3.161–80.
, and D
. 1977. Face-to-face interaction: Research, meth-
ods and theory. Hillsdale, NJ: Lawrence Erlbaum.
E., and S
. 1996. Interactional units in conversation:
Syntactic, intonational, and pragmatic resources for the management of turns. Interac-
tion and grammar, ed. by Emanuel A. Schegloff and Sandra A. Thompson, 135–84.
Cambridge: Cambridge University Press.
; and S
. 2003. Social interaction and
grammar. The new psychology of language: Cognitive and functional approaches to
language structure, ed. by Michael Tomasello, 119–44. Mahwah, NJ: Lawrence Erl-
E. 1995. Constructions: A construction grammar approach to argument
structure. Chicago: University of Chicago Press.
. 1979. The interactive construction of a sentence in a natural conversa-
tion. Everyday language: Studies in ethnomethodology, ed. by George Psathas, 97–121.
New York: Irvington.
. 1981. Conversational organization: Interaction between speakers and
hearers. New York: Academic Press.
LANGUAGE, VOLUME 82, NUMBER 3 (2006)534
. 2006. Human sociality as mutual orientation in a rich interactive envi-
ronment: Multimodal utterances and pointing in aphasia. Roots of human sociality:
Culture, cognition and interaction, ed. by N. J. Enfield and Stephen C. Levinson,
97–125. London: Berg.
. 1980. Spoken word recognition processes and the gating paradigm.
Perception & Psychophysics 28.267–83.
, and C
. 1996. Using prosody to predict the end of sen-
tences in English and French: Normal and brain-damaged subjects. Language and
Cognitive Processes 11.107–34.
. 2005. Transcription of Dutch intonation. Prosodic typology: The
phonology of intonation and phrasing, ed. by Sun-Ah Jun, 118– 45. Oxford: Oxford
. 2005. Joint turn construction through language and the body: Notes on
embodiment in coordinated participation in situated activities. Semiotica 156.21–53.
. 1984. Garfinkel and ethnomethodology. Cambridge: Polity Press.
. 1998. Emergent grammar. The new psychology of language: Cognitive and
functional approaches to language structure, ed. by Michael Tomasello, 155–75. Mah-
wah, NJ: Lawrence Erlbaum.
. 1973. A case of precision timing in ordinary conversation: Overlapped
tag-positioned address terms in closing sequences. Semiotica 9.47–96.
. 1967. Some functions of gaze-direction in social interaction. Acta Psychol-
. 2004. Gesture: Visible action as utterance. Cambridge: Cambridge Univer-
. 1994. Information structure and sentence form: Topic, focus and the
mental representations of discourse referents/grammatical relations. Cambridge: Cam-
bridge University Press.
W. 1987. Foundations of cognitive grammar, vol. 1: Theoretical
prerequisites. Stanford, CA: Stanford University Press.
C. 1983. Pragmatics. Cambridge: Cambridge University Press.
; and B
. 1986. Towards a phonology of conversation:
Turn-taking in Tyneside English. Journal of Linguistics 22.411–37.
; and M
. 1985. Phonetic aspects of turn delimination
in London Jamaican. Journal of Pragmatics 9.309–30.
K. 1989. Japanese conversation: Self contextualization through structure
and interactional management. Norwood, NJ: Ablex.
. 1992. Hand and mind. Chicago: Chicago University Press.
J. 1983. The origin of sound patterns in vocal tract constraints. The production
of speech, ed. by Peter F. MacNeilage, 189–216. New York: Springer.
. 1974. Contrapuntal conversations in an Antiguan village. Language in its
social setting, ed. by Richard Bauman and Joel Sherzer, 110– 24. Cambridge: Cam-
bridge University Press.
. 2005. On sustaining vs. withdrawing gaze in face-to-face interaction.
Paper presented at 91st annual convention of the National Communication Association,
Boston, MA, November 2005.
; and G
. 1974. A simplest systematics
for the organization of turn-taking for conversation. Language 50.696–735.
. 1983. The role of intonation as a cue to turn taking in conversation.
Journal of Phonetics 11.243–57.
A. 1982. Discourse as an interactional achievement: Some uses of
‘uh huh’ and other things that come between sentences. Analyzing discourse: Text and
talk, ed. by Deborah Tannen, 71– 93. Washington, DC: Georgetown University Press.
A. 1996. Turn organization: One intersection of grammar and interac-
tion. Interaction and grammar, ed. by Elinor Ochs, Emanuel A. Schegloff, and Sandra
A. Thompson, 52–133. Cambridge: Cambridge University Press.
A. 2000. Overlapping talk and the organization of turn-taking for
conversation. Language in Society 29.1–63.
E. 1948. A mathematical theory of communication. Bell System Technical
Journal 27.379–423, 623– 56.
PROJECTING THE END OF A SPEAKER’S TURN 535
. 2001. Conversational turn-taking in a Caribbean English Creole. Journal of
. 1985. Cross cultural communication. Handbook of discourse analysis,
ed. by Teun A. Van Dijk, 203–15. London: Academic Press.
A., and P
. 2001. Transitivity, clause structure and argu-
ment structure: Evidence from conversation. Frequency and the emergence of linguistic
structure, ed. by Joan L. Bybee and Paul J. Hopper, 27–60. Amsterdam: John Benja-
, and G
. 2000. Syntactic structure assembly in human parsing:
A computational model based on competitive inhibition and a lexicalist grammar. Cog-
, and S
. 1996. Ending up in Ulster: Prosody and turn-taking in English
dialects. In Couper-Kuhlen & Selting 1996b, 101–30.
. 1988. The semantics of grammar. Amsterdam: John Benjamins.
H. 1970. On getting a word in edgewise. Chicago Linguistic Society
Max Planck Institute for Psycholinguistics [Received 2 May 2005;
P.O. Box 310 accepted 21 October 2005]
NL-6500 AH Nijmegen