Umm... With Transformers? Insights from Filled Pause Use across Four Slavic Parliaments

Technical Summary: Insights from Filled Pause Use across Four Slavic Parliaments

Problem Statement
Filled pauses (FPs), such as "uh" and "um," are universal features of spontaneous speech, yet empirical research into their production has been constrained by reliance on small, single-language corpora. This limitation restricts statistical power and the generalizability of findings across linguistic and cultural contexts. Furthermore, existing literature often conflates stable speaker traits with state-level variations within utterances, making it difficult to distinguish between habitual speaking styles and moment-to-moment planning demands. This study addresses these gaps by analyzing approximately 4,000 hours of parliamentary speech across four related Slavic languages (Croatian, Czech, Polish, and Serbian) to investigate the predictors of FP rates in a large-scale, linguistically diverse setting.

Methodology
The research utilizes the ParlaSpeech dataset, comprising 6,000 hours of parliamentary speech, filtered to include 1,001,787 utterances from 1,561 speakers (regular members of parliament) with utterances of at least 3 seconds and 10 words.

• Automatic Detection: Filled pauses were identified using a transformer-based model (wav2vec2-bert) trained on Slovenian data and tested across the four target languages, achieving an event-level F1 score of approximately 0.92.
• Feature Extraction: Speaker metadata (gender, age, political orientation, power status) were sourced from official records. Sentiment was predicted using XLM-R-ParlaSent, and speech rate was calculated as syllabic vowel count divided by audio duration.
• Statistical Modeling: The study employs Generalised Estimating Equations (GEE) with a Negative Binomial distribution to model FP counts, using log(audio duration) as an offset to express results as rates.
• Decomposition Strategy: A crucial methodological component is the application of Mundlak-corrected models. This approach decomposes time-varying predictors into two components:
  1. Between-speaker effects: Stable tendencies of a speaker (e.g., a speaker's average age or habitual sentiment).
  2. Within-speaker effects: Deviations from a speaker's own baseline during specific utterances (e.g., a speaker speaking more positively than usual in a specific session).
     This decomposition allows the authors to distinguish trait-level associations from state-level variations. Ten models were fitted: baseline and Mundlak-corrected specifications, estimated both globally (pooled) and separately for each parliament.

Key Results
The analysis yields distinct findings regarding established and novel predictors of FP rates:

• Gender: Contrary to much prior literature reporting higher FP rates in men, this study finds that globally, male speakers produce significantly fewer FPs than female speakers (Incidence Rate Ratio [IRR] = 0.636). This effect is robust and driven entirely by South Slavic parliaments (Croatia and Serbia), where the reduction is substantial. In West Slavic parliaments (Czech and Poland), gender differences were non-significant.
• Age: A negative association between age and FP rate was confirmed. Globally, each additional decade of age corresponds to a 13.8% reduction in FPs. This contrasts with some conversational corpora that report increased FPs with age, reinforcing the domain-dependence of age effects.
• Speech Rate: A strong negative correlation exists between speech rate and FP rate. The within-speaker component is the strongest predictor in the dataset, indicating that moment-to-moment planning demands (slowing down and hesitating) drive FP production more than a speaker's habitual pace.
• Sentiment: A consistent positive association was found between sentiment and FP rate. Speakers use more FPs when expressing more positive sentiment. This effect is stable within speakers across all four parliaments, suggesting that positive communicative demands increase hesitation, regardless of the speaker's baseline style.
• Political Orientation: No reliable global association was found between a speaker's habitual ideological position and FP production. However, within-speaker analysis revealed that in Croatia and Poland, sessions where a speaker's party affiliation was more right-leaning than their usual baseline were associated with higher FP rates. This effect was absent in Czechia and Serbia.
• Power Status: Opposition speakers globally tend to produce fewer FPs than ruling coalition members (IRR = 0.792). This effect was significant in Czechia and Croatia. Notably, Serbia showed a borderline positive trend (opposition speaking with more FPs), suggesting unique political dynamics in that context.

Significance and Contributions
The paper claims several significant contributions to the field of computational paralinguistics:

1. Scale and Generalizability: By analyzing nearly 4,000 hours of speech across four languages, the study demonstrates that many effects reported in smaller, single-language studies (particularly regarding gender and age) are not universal but are instead language-specific or domain-dependent.
2. Methodological Rigor: The application of Mundlak-corrected GEE models provides a robust framework for disentangling stable speaker traits from utterance-level variation. The results highlight that the within-speaker component (state-level variation) is often a stronger predictor of FP production than between-speaker traits, particularly for speech rate and sentiment.
3. Novel Predictors: The study establishes sentiment as a consistent predictor of FP rates, a relationship rarely explored in prior quantitative work. It also provides the first large-scale quantitative evidence regarding the impact of political orientation and power status on disfluency, revealing complex, parliament-specific patterns rather than simple linear trends.
4. Domain Specificity: The authors emphasize that their findings are specific to the parliamentary domain. The reversal of gender effects in South Slavic parliaments and the specific modulation of FP rates by power status suggest that formal, task-oriented speech genres may neutralize or invert patterns observed in informal conversation.

The paper concludes that large, transformer-annotated, cross-lingual corpora are essential for understanding FP production and that caution is required when generalizing findings from small-scale studies to broader linguistic or cultural contexts.