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

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies, Data Science & Analytics · Depth: Expert, long

Summary

An analysis of approximately 4,000 hours of parliamentary speech across Croatian, Czech, Polish, and Serbian parliaments investigated the use of filled pauses (FPs). Utilizing transformer-based automatic detection for FP identification and Generalised Estimating Equations (GEE) with Mundlak correction, the study distinguished between-speaker and within-speaker effects on FP rates. It confirmed a negative association between age and speech rate with FP occurrence. Notably, gender effects were found to be language-specific, with male speakers producing significantly fewer FPs than female speakers in Croatian and Serbian parliaments, a reversal of most prior literature. Novel analyses revealed a consistent positive association between sentiment and FP rate, meaning more positive language correlates with more FPs. Political orientation and power status showed more fragmented, parliament-specific effects, though opposition speakers generally exhibited lower FP rates than governing coalition speakers.

Key takeaway

For NLP Engineers developing speech analysis tools for cross-lingual applications, you should recognize that filled pause (FP) patterns are not universal. Your models must account for language- and culture-specific variations, especially regarding gender effects, which can reverse across different Slavic languages. Consider incorporating within-speaker analysis to distinguish habitual traits from situational speech variations, improving the accuracy and generalizability of your disfluency detection and interpretation.

Key insights

Filled pause patterns are highly context-dependent, varying significantly across languages and discourse domains.

Principles

• FP rates decrease with age and faster speech.
• Gender effects on FPs are language- and context-specific.
• Sentiment positively correlates with FP rate.

Method

Transformer-based models detect filled pauses at scale. Generalised Estimating Equations (GEE) with Mundlak correction separate within-speaker from between-speaker effects on FP rates, using Negative Binomial regression.

In practice

• Use transformer models for large-scale disfluency analysis.
• Account for linguistic and cultural context in speech analysis.

Topics

• Filled Pauses
• Transformer Models
• Computational Paralinguistics
• Slavic Languages
• Parliamentary Speech
• Speaker Traits

Best for: AI Scientist, Research Scientist, NLP Engineer

Related on AIssential

• PolySpeech-100: A Large-Scale Benchmark for Speech Understanding Across 100+ Languages and Dialects — PolySpeech-100 reveals E2E Speech-LLMs excel on dialects, but open-source models struggle with low-resource languages, and CoT prompting can degrade performance.
• NaturalFlow: Reducing Disruptive Pauses for Natural Speech Flow in Simultaneous Speech-to-Speech Translation — Simultaneous S2ST can achieve natural speech flow by optimizing for fluency and translation quality concurrently, avoiding disruptive pauses.
• Phonological Subspace Collapse Is Aetiology-Specific and Cross-Lingually Stable: Evidence from 3,374 Speakers — Phonological subspace analysis offers a robust, training-free, and cross-lingually stable method for dysarthria characterization.
• Paza: Introducing automatic speech recognition benchmarks and models for low resource languages — Human-centered design and evaluation are crucial for effective speech models in low-resource, real-world environments.
• Covertly improving intelligibility with data-driven adaptations of speech timing — Targeted speech rate adjustments, not global slowing, covertly improve intelligibility across diverse listener groups.
• L-Proto: Language-Aware Episodic Prototypical Training for Multilingual Speaker Verification — L-Proto improves multilingual speaker verification by using language-consistent episodic training to disentangle speaker identity from linguistic characteristics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.