TL;DR
This paper explores how speakers unconsciously adjust their speech rhythm and pitch (prosody) when working together on a task. They introduce and apply novel methods to quantify this "prosodic alignment," finding that speakers do indeed synchronize their prosody, particularly their intensity, and that this alignment might be stronger when speakers are unfamiliar with each other.
What Problem Does This Paper Address?
The Speech Accommodation Theory suggests that people unconsciously adapt their speech to their conversational partners to build rapport. While this phenomenon has been studied for lexical and syntactic alignment, its presence and measurement in prosody (pitch, rhythm, intensity) during cooperative tasks remain less understood. Existing methods for measuring prosodic alignment are often qualitative or lack robust ways to distinguish genuine alignment from random coincidences. This paper aims to bridge this gap by developing and applying quantitative measures to assess prosodic convergence and synchrony in task-based conversations, and to investigate factors like speaker familiarity and task success.
Methods
The authors employ a dynamic approach to measure prosodic alignment, moving beyond static comparisons. Key methods include:
- Talkspurt Creation: Speech segments were segmented into "talkspurts" by bridging short silences and speech events to create workable units.
- Feature Processing: Prosodic features like Log Fo (pitch) and intensity were extracted and smoothed using moving windows. To handle the discontinuous nature of speech, these features were transformed into continuous streams and then normalized using z-scores.
- Convergence Measurement: Two methods were used:
- Correlation with Time: Measuring the Pearson correlation between the difference in speakers' feature values and time. A negative correlation suggests convergence.
- Half-Conversation Comparison: Comparing the average feature values in the first and second halves of a conversation. A positive difference (first half minus second half) indicates convergence.
- Synchrony Measurement: A windowed cross-correlation (wcc) procedure, adapted from visual behavior research, was used. This method allows for the analysis of dynamic alignment and potential time lags between speakers' prosodic features. It involves calculating correlations over moving windows, allowing for shifts (lags) between the two signals.
- Distinguishing Genuine Alignment: To address the issue of random coincidence, the authors generated "pseudointeractions." This involved pairing a speaker's real speech with shuffled speech from another speaker (or a different part of the same speaker's speech) to create a baseline. Genuine alignment should be stronger than alignment observed in these pseudointeractions.
Datasets
The study utilizes the HCRC Map Task Corpus. This corpus contains Scottish English spoken task-based dyadic conversations. The researchers selected 31 "familiar" (FAM) and 32 "unfamiliar" (UNFAM) conversations from the "no-eye-contact" condition to focus on vocal alignment. Task success was measured by the deviation of the follower's drawn route from the giver's intended route.
Key Insights and Ideas
- Prosodic Alignment is Real: The study provides evidence that speakers do exhibit prosodic convergence and synchrony during cooperative task-based conversations.
- Intensity is Key: Alignment tendencies were strongest for intensity (both absolute and relative) compared to pitch (Log Fo).
- Familiarity Matters (Potentially): There were indications that unfamiliar speakers might show stronger convergence than familiar speakers, possibly because they need to establish rapport. However, this finding requires further investigation.
- Pseudo-Interaction Baseline is Crucial: The authors highlight the importance of using methods like pseudointeractions to ensure that observed alignment is not just a statistical artifact but reflects genuine speaker-specific adaptation.
- Dynamic Measurement is Necessary: The use of windowed cross-correlation demonstrates the value of dynamic approaches to capture the temporal nature of prosodic alignment.
Key Results
- Convergence: Low levels of convergence were observed overall. Intensity showed signs of convergence, with tendencies for stronger convergence in unfamiliar pairs.
- Synchrony: Generally low synchrony was found. Intensity showed more synchrony than pitch. The difference between familiar and unfamiliar pairs was not significant in most cases.
- Pseudo-Interaction Comparison: Most pseudointeractions yielded significantly lower synchrony levels than real interactions, supporting the idea of speaker-specific alignment.
- Task Success: A weak positive correlation was found between alignment of relative intensity and task success. A multiple regression analysis indicated that alignment measures explained about 13% of the variance in task success.
Limitations and Future Work
- Limited Significance: While tendencies were observed, many results did not reach statistical significance, making definitive conclusions challenging.
- Familiarity Effect Nuance: The observed effect of familiarity requires further investigation to be fully understood.
- Task Success Correlation: The correlation with task success was weak, suggesting that prosodic alignment is only one factor among many influencing task outcomes.
- Future Directions: The authors suggest concentrating on the dynamics of alignment and explicitly analyzing leading or following behaviors (time lags) to gain deeper insights into social dynamics. Further research is also needed to refine the measurement and representation of these dynamic alignment processes.
Summary updated