robson

Conversations Between Friends & Strangers

Nicole Robson

1. Abstract

Despite its anarchic and spontaneous nature, natural conversation is remarkably smooth with between-speaker silences far smaller (< 200ms) than the time required to form and begin a new word or phrase. This virtuoso coordination indicates that we must predict where our conversational partner is likely to finish their utterance and prosodic, lexical, syntactic and visual cues may all play a part in this prediction. This project explores the possibility that familiarity with a conversational partner informs our ability to predict their pauses and ends of phrases by comparing the timing of turn-exchanges and backchannels in dyadic conversations between friends and strangers.

2. Introduction

Conversation is the primary way that we use language and how children learn to speak. The ability to coordinate our speech in relation to others is learned at a very early age and from two years, children learn to curb simultaneous speech and take on a listener role [14]. Conversational turns are unfixed in length and adapt to the open-ended and spontaneous character of natural language. Yet most often there are only very short silences between speakers and little unintentional overlap. It is rare for two speakers to begin an utterance at the very same time [1].

The pauses and gaps found at turn-exchanges are often smaller than 200ms, as shown in this transcription of a typical telephone conversation between friends [9].

Mat: ‘lo Redcah five o’six one?

Ver: [0.15] Hello Mathew is yer mum the:hr love.

Mat: [0.13] Uh no she’s, gone (up) t’town,h

Ver: [0.24] Al:right uh will yih tell’er Antie Vera rahn:g then.

Mat: [0.03] Yeh.

Ver: [0.13] Okay. She’s alright is she.

Mat: [0.10] Yeh,h

Ver: [0.07] Okay. Right. Bye bye luv,

Mat: [0.02] Tara, .h

(End call)

Our language production system is slow and requires 600ms from word conception to output [11], with multiple words and phrases requiring even longer [8]. Therefore, if there are only tiny pauses between turn-exchanges of 200ms, we must anticipate turn-endings very effectively. Magyari et al. found EEG data to show even earlier neural activity before turn-endings of up to 1800ms [4].

Turn-end predictions might be facilitated by a range of cues such as a prosodic rise or fall at the phrase end [12] and visual cues such as gaze [5]. However, these provide only binary information at the very end of a turn as to whether it is about to end or not and happen too late to account for the tiny offsets when compared to response times. As a sentence unfolds the probability of it continuing in any certain direction narrows and anticipated syntax or meaning might provide earlier and finer grained information about a turn-ending.

This project is motivated by the evidence found for turn-end prediction and compares the coordination of spoken interactions between friends and strangers. I hypothesise that conversations between friends are characterised by closely coordinated turn-exchanges and backchannels, with shorter offsets than between strangers. It is possible that we learn the individual characteristics and style of our friends' speech, such as rate, prosody, syllable articulation. Their world view or character might also inform our predictions for the outcome of a phrase. This prior knowledge sets up expectations and might allow us to better predict where their pauses and ends of phrases will fall.

To test this theory, I will analyse two features of turn-taking in conversation:

1. Turn-exchanges: The transition from one speaker to the next and specifically where the next speaker expresses an intent to take the floor
2. Backchannels: Short utterances offered in support of the speaker with the floor, that do not signal an intention to take the floor. Instead they communicate empathy or understanding of a subject, e.g. 'mmm', 'yeah', 'okay'.

In this study, friends are defined simply as participants with some prior experience conversing with each other and may be colleagues, friends or family members. The level of experience/ familiarity with each other is not specified. Strangers are simply classified as participants who had not met or conversed prior to the day of recording.

3. Related Work

The methodology for this project is modelled on studies by Heldner [7] and Lunsford [6] that analyse pauses, gaps and overlaps in conversation. It also builds on a classification system for turn-exchanges and backchannels proposed by Beattie [13].

4. Dataset

The conversations analysed in this study are taken from the archive of my project The Conversation Booth: 100 conversations recorded at 7 different events around London in 2017-18. The Conversation Booth is a pop-up sound installation that invites visitors to record themselves discussing a particular topic or question. Their responses are automatically archived by a computer program and recomposed as part of a living speech collage heard on wireless headphones at the installation and later at the project website.

Figure 1. Visitors to the The Conversation Booth.

The installation locations and the topic of conversation discussed for conversations used in this study are listed below:

The booth is acoustically isolated but not soundproofed and participants sit beside each other, with both voices captured by one microphone positioned slightly behind their bodies, so as not to visually distract or influence their behaviour. The Conversation Booth is not a controlled experimental environment but it does encourage focused and natural conversations. There are a number of confounding variables that may affect the results but for me it was important to study conversation with all of its spontaneity and individual quirks. The purpose of this project is to see what patterns emerge in spite of this.

I selected 16 conversations from the archive, 8 between friends and 8 between strangers. The first three minutes of each conversation was extracted for analysis.

5. Method

Each audio recording was annotated and coded in TextGrids using the software Praat. These produced labelled numerical datapoints that could be exported as .csv files for analysis and visualisation.

i. Annotation

As a first step I removed background noise where necessary and automatically annotated sound and silence to a TextGrid using the built-in Praat function, adjusting the parameters to best suit each recording. These computed values were used as a framework to inform the manual annotation of turns and double checked for accuracy.

The first speaker heard in each recording took the label 'S1' and their conversational partner 'S2'. The beginning and end of each turn was annotated and labelled with the correct speaker, while turn-exchanges- the periods of silence between speaker turns- were labelled as 'silent'. Where turns overlapped, the period of simultaneous speech was labelled 'S12'.

On a separate tier I annotated every instance of a backchannel, labelled with the speaker involved, the silences before or after it and any partially overlapping speech by the speaker with the floor.

Figure 2. An example TextGrid annotation in Praat with 3 tiers for Turns, Codes & Backchannels

ii. Coding

I coded every turn-exchange and backchannel based on the system outlined by Beattie [13]. The intention of a speaker to take the floor is key for distinguishing turn-exchanges from backchannels.

Figure 3. Beattie rationale and system for coding turn-exchanges and backchannels.

Turn-exchanges were coded as:

Smooth switch (S): Silence between turns, S1 utterance complete

Pause Interruption (PI): Silence between turns, S1 utterance incomplete

Overlap (O): The beginning of S2's speech overlaps the end of S1's turn, which is otherwise complete

Interruption (I): The beginning of S2's speech overlaps the end of S1's turn

Butting-in (BI): An unsuccessful interruption, S2 overlaps but doesn't gain the floor

After hearing highly coordinated backchannelling in the dataset, I developed the codes in order to describe these interactions in greater detail. Some backchannels are clearly not intended to synchronise or coordinate with the main speaker, they occur mid-phrase and are completely overlapped. Other backchannels fall intentionally at the end of a phrase or within a pause. In these cases I believe that S2 intends to coordinate their backchannel to be heard independently, with little or no overlap.

Backchannels were therefore coded as:

BC = Backchannel, no overlap, articulated by silence at start and end

BC_OC = Backchannel, completely overlapping speech (within phrase, no attempt to sync)

BC_OP_S = Backchannel, partial overlap at the start (at phrase end, attempts to sync)

BC_OP_E = Backchannel, partial overlap at the end (at phrase end, attempts to sync)

6. Results

The TextGrid files were converted to tables and saved as .csv, which allowed me to organise and summarise the data.

Table 1 shows that a similar number of turn-exchanges of most types are found for the groups of friends and strangers. The only notable difference is the frequency of 'Butting-In' interactions which is almost twice as many (9) for friends.

Table 1. Overview of results showing the number of different turn-exchange types.

Table 2 shows that almost twice the number of BC_OC types occurred in the friends group, while the strangers exhibited twice the number of BC_OP_E types. It is possible these partial overlaps represent failed BC backchannels.

Table 2. Overview of results showing the number of different backchannel types.

Figure 4 visualises the turn distributions of both groups and shows that friends shared the floor more evenly than strangers, while strangers tended toward an interview-like format, with one speaker dominating the conversation. There is a greater overlap of speech at turn-exchanges for friends, compared with strangers.

Figure 4. Plots showing the distribution of turns in conversations between friends and strangers.

Figure 5 shows frequency distribution plots of turn-exchange offsets for friends and strangers. For both groups these fall most frequently in the range of 100-200ms and the distributions are very similar. Among friends however, there is a greater frequency of overlapping speech, with longer overlaps of up to 2 seconds.

Figure 5. Frequency distributions of turn-exchange offsets

Figure 6 shows greater difference in the frequency of offset durations for backchannels. In these case speech is most often offset by 0-100ms among friends, with comparatively few overlaps. For strangers, offsets are most frequently in the range of 200-300ms and with more frequent overlaps (30, compared with 22). The coordination of their backchannels in relation to speech is less accurate than friends and less accurate than their coordination of turn-exchanges.

[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

Figure 6. Frequency distributions of backchannel offsets

7. Discussion

Strangers conform more closely to the 'one speaker at a time' paradigm and perhaps this is considered most polite in a first interaction. Interruptions, overlapping turn-exchanges and butting-in were found to be more common among friends. Interruptions are traditionally viewed as a sign of social dominance or of discomfort, but may also signal heightened involvement in the conversation and joint enthusiasm [13]. All of the conversations used in this study could be described as supportive and free of antagonism, therefore I believe that interruptions and overlaps result from the heightened collaborative spirit of conversation between friends.

The length of turn-exchange overlaps among friends may also be significant. Magyari et al. found early anticipation of turn-endings of up to 1800ms when the lexical content of a phrase was classified as 'predictable', rather than 'unpredictable' [4]. Twice as many turn-exchanges among friends had an overlap of >1300ms and this may indicate that friends predict the outcome of their partners turn earlier, or simply that they feel more free to act on that prediction.

Backchannelling was used more by friends and with greater precision at phrase endings. The study therefore presents preliminary evidence for the smoother coordination of backchannels among this group. This may be achieved by the collaboration of both speakers, with S1 coordinating the start of their backchannel with the tail end of S2's phrase and S2 coordinating the beginning of their next utterance with the tail end of the S1's backchannel. Alternatively the partner without the floor may predict the length of an upcoming pause and insert a backchannel of the correct length accordingly. Conversational experience and an understanding of their partners accent, character or style might enable this close coordination.

The Conversation Booth dataset used for this project does not have any control over the cultural background of a speaker and includes many conversations between members of migrant communities in London. In most cases, English is their first language but this may not always be the case and the presence of less experienced English speakers may have a confounding effect on some of the results; longer pauses may be required to facilitate understanding or to form a new sentence for example. However, London's population is fundamentally multi-cultural and that diversity should be celebrated. To best understand how spoken interactions operate in the real world it seems sensible to recruit participants from across a broad demographic. The fact that patterns emerge in this project regardless of possible cultural and language biases supports this belief.

8. Conclusions

This project explored the possibility that familiarity with a conversational partner allows us to predict features of their speech and compared the timing of turn-exchanges and backchannels in dyadic conversations between friends and strangers. The initial findings are promising but a much larger sample size is needed to better confirm their validity.

While the timing of turn-exchanges between friends and strangers are very similar, friends tend to overlap, interrupt and but-in with greater frequency. This is perceived not as a struggle for dominance but as characteristic of enthusiasm and mutual involvement involvement in the conversation. I hypothesised that speech between friends would show tighter coordination of turn-exchanges and this was not the case. However, the placement of backchannels do show greater coordination among friends and this is an unexpected finding, not found in the literature and warranting further investigation.

9. Further Work

A clear weakness in this project is the small sample size and length of audio files which should both be extended for further analyses. The level of manual annotation required to gather accurate data for this project was prohibitive and in future, a more computational approach to annotation would be necessary. Separate audio streams for each speaker would enable automatic annotation of turns but these do not exist for The Conversation Booth dataset. Spectral methods might enable automatic annotation for some of the dyadic conversations where the voices are substantially different, but similar voices (a feature among friends and family) would still be problematic and require manual annotation.

By increasing the length of extracts, or taking one extract from the beginning of a conversation and the other from the end, it would be interesting to see whether the coordination of turn-exchanges and backchannels improves over the course of a conversation, especially among strangers. There may also be a relationship between offsets and speech rate. Wilson & Wilson [2] and Large [10] suggest that the placement of turns is dictated by rhythmic entrainment of neural oscillators in the brain. It may be that the smaller offsets found among friends are due to a faster rate of speech, rather than familiarity or experience conversing. Similarly, the characteristics that I have highlighted in this study may have stronger links to perceived levels of affiliation, dominance or formality rather than the simplistic division of friends and strangers. It would be worth regrouping conversations along these different lines to explore this idea or running a separate listening test to code conversations within these categories.

10. References

H. Sacks, E. A. Schegloff, and G. Jefferson, “A Simplest Systematics for the Organization of Turn-Taking for Conversation,” Language, vol. 50, no. 4, p. 696, Dec. 1974.

M. Wilson and T. P. Wilson, “An oscillator model of the timing of turn-taking,” Psychonomic Bulletin & Review, vol. 12, no. 6, pp. 957–968, Dec. 2005.

S. Benus, “Are We `in Sync’: Turn-Taking in Collaborative Dialogues,” p. 4.

L. Magyari, M. C. M. Bastiaansen, J. P. de Ruiter, and S. C. Levinson, “Early Anticipation Lies behind the Speed of Response in Conversation,” Journal of Cognitive Neuroscience, vol. 26, no. 11, pp. 2530–2539, Nov. 2014.

F. Rossano, P. Brown, and S. C. Levinson, “Gaze, questioning, and culture,” in Conversation Analysis, J. Sidnell, Ed. Cambridge: Cambridge University Press, 2009, pp. 187–249.

R. Lunsford, P. A. Heeman, and E. Rennie, “Measuring Turn-Taking Offsets in Human-Human Dialogues,” presented at the Interspeech 2016, 2016, pp. 2895–2899.

M. Heldner and J. Edlund, “Pauses, gaps and overlaps in conversations,” Journal of Phonetics, vol. 38, no. 4, pp. 555–568, Oct. 2010.

T. T. Schnur, A. Costa, and A. Caramazza, “Planning at the Phonological Level during Sentence Production,” Journal of Psycholinguistic Research, vol. 35, no. 2, pp. 189–213, Mar. 2006.

J.-P. de Ruiter, H. Mitterer, and N. J. Enfield, “Projecting the End of a Speaker’s Turn: A Cognitive Cornerstone of Conversation,” Language, vol. 82, no. 3, pp. 515–535, 2006.

E. W. Large and M. R. Jones, “The dynamics of attending: How people track time-varying events,” Psychological Review, vol. 106, no. 1, pp. 119–159, Jan. 1999.

P. Indefrey, “The Spatial and Temporal Signatures of Word Production Components: A Critical Update,” Frontiers in Psychology, vol. 2, 2011.

E. A. Schegloff, “Turn organization: one intersection of grammar and interaction,” in Interaction and grammar, E. Ochs, E. A. Schegloff, and S. A. Thompson, Eds. Cambridge: Cambridge University Press, 1996, pp. 52–133.

G. W. Beattie, “Turn-taking and interruption in political interviews: Margaret Thatcher and Jim Callaghan compared and contrasted,” Semiotica, vol. 39, no. 1–2, 1982.

D. N. Stern, J. Jaffe, B. Beebe, and S. L. Bennett, “Vocalising in Unison and in Alternation: Two modes of communication within the mother-infant dyad,” Annals of the New York Academy of Sciences, vol. 263, no. 1 Developmental, pp. 89–100, Sep. 1975.

Page updated

Google Sites

Report abuse