pith. sign in

arxiv: 2604.13620 · v1 · submitted 2026-04-15 · 💻 cs.CL · cs.AI

Syn-TurnTurk: A Synthetic Dataset for Turn-Taking Prediction in Turkish Dialogues

Ahmet Tu\u{g}rul Bayrak , Mustafa Serta\c{c} T\"urkel , Fatma Nur Korkmaz This is my paper

Pith reviewed 2026-05-10 13:27 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords synthetic datasetturn-taking predictionTurkish dialoguesdialogue timingLLM generated dataBI-LSTM modelvoice chatbotsnatural interaction
0
0 comments X

The pith

A synthetic LLM-generated dataset of Turkish dialogues enables models to predict turn-taking at 0.839 accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper presents Syn-TurnTurk, a dataset of artificial Turkish conversations created with large language models to include features like overlaps and strategic pauses that real people use. The authors test multiple machine learning models on the dataset to predict when one speaker should yield to another. Top performing models such as bidirectional LSTM and an ensemble of logistic regression with random forest reach an accuracy of 0.839 and an AUC of 0.910. These results indicate that the synthetic data helps models learn the timing signals in Turkish speech, which could improve voice interfaces so they interrupt less and flow more naturally. The work targets the shortage of suitable real data for this language and task.

Core claim

The paper introduces Syn-TurnTurk, a synthetic Turkish dialogue dataset generated using various Qwen Large Language Models to mirror real-life verbal exchanges, including overlaps and strategic silences. Evaluation using traditional and deep learning architectures shows that BI-LSTM and Ensemble (LR+RF) methods achieve high accuracy of 0.839 and AUC scores of 0.910. This demonstrates that the synthetic dataset positively affects models' ability to understand linguistic cues, supporting more natural human-machine interaction in Turkish.

What carries the argument

Syn-TurnTurk, the LLM-generated synthetic dataset that replicates turn-taking behaviors such as overlaps and silences in Turkish dialogues for training prediction models.

If this is right

  • Voice-based chatbots for Turkish can reduce interruptions by using models trained to predict turns from the synthetic data.
  • BI-LSTM and ensemble models learn effectively from the included overlaps and silences.
  • The dataset fills a gap for Turkish, enabling progress where real data is unavailable.
  • High AUC scores indicate strong performance in distinguishing turn-taking events.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Since no real data comparison is reported, the transfer to live conversations remains untested.
  • The method of generating dialogues with LLMs may extend to modeling other dialogue phenomena in Turkish.
  • Future datasets could combine synthetic and real elements to strengthen the approach.

Load-bearing premise

That the turn-taking patterns, overlaps, and silences in the LLM-generated dialogues match those in real spoken Turkish without any validation against human recordings.

What would settle it

Compare the performance of models trained and tested on Syn-TurnTurk against the same models tested on a collection of real human Turkish dialogues; a large drop in accuracy would indicate the synthetic data does not capture the true patterns.

Figures

Figures reproduced from arXiv: 2604.13620 by Ahmet Tu\u{g}rul Bayrak, Fatma Nur Korkmaz, Mustafa Serta\c{c} T\"urkel.

Figure 1
Figure 1. Figure 1: FTO distribution histogram IV. TURN PREDICTION MODELS To evaluate the effectiveness of the generated dataset, sev￾eral classification models were implemented, from traditional machine learning algorithms to advanced deep learning meth￾ods. Specifically, we utilized Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and Bidirectional Long Short-Term Memory (BI-LSTM) architectures. In each tur… view at source ↗
read the original abstract

Managing natural dialogue timing is a significant challenge for voice-based chatbots. Most current systems usually rely on simple silence detection, which often fails because human speech patterns involve irregular pauses. This causes bots to interrupt users, breaking the conversational flow. This problem is even more severe for languages like Turkish, which lack high-quality datasets for turn-taking prediction. This paper introduces Syn-TurnTurk, a synthetic Turkish dialogue dataset generated using various Qwen Large Language Models (LLMs) to mirror real-life verbal exchanges, including overlaps and strategic silences. We evaluated the dataset using several traditional and deep learning architectures. The results show that advanced models, particularly BI-LSTM and Ensemble (LR+RF) methods, achieve high accuracy (0.839) and AUC scores (0.910). These findings demonstrate that our synthetic dataset can have a positive affect for models understand linguistic cues, allowing for more natural human-machine interaction in Turkish.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces Syn-TurnTurk, a synthetic Turkish dialogue dataset generated using Qwen LLMs to include overlaps and strategic silences for turn-taking prediction. It evaluates traditional ML and deep learning models (e.g., BI-LSTM achieving 0.839 accuracy and 0.910 AUC) on this dataset and claims the resource enables models to better understand linguistic cues for natural human-machine interaction in Turkish.

Significance. A validated synthetic dataset for turn-taking in Turkish would address a clear resource gap for low-resource languages and could improve voice chatbot naturalness beyond simple silence detection. The reported internal performance of BI-LSTM and ensemble models is promising for synthetic data, but the absence of any external validation or baseline comparisons means the central claim of positive effect on real linguistic-cue understanding is not yet substantiated.

major comments (3)
  1. Abstract: the assertion that the dataset 'mirrors real-life verbal exchanges, including overlaps and strategic silences' is unsupported because no quantitative validation (e.g., pause-duration histograms or overlap statistics) against any human-recorded Turkish corpus is reported; this directly undermines the claim that models trained on it learn genuine linguistic cues rather than artifacts of the generation process.
  2. Abstract / Results: all performance figures (BI-LSTM 0.839 accuracy, 0.910 AUC; Ensemble LR+RF) are obtained from models trained and tested on the same synthetic data with no held-out real Turkish test set and no comparison to the silence-detection baseline explicitly mentioned in the introduction; this circularity makes the generalization claim load-bearing and untested.
  3. Data generation description: the manuscript provides no details on prompting strategies, temperature settings, or post-processing used to elicit overlaps and strategic silences from Qwen models, nor any statistical fidelity checks, leaving the core premise that the synthetic data reproduces human Turkish turn-taking patterns unverified.
minor comments (2)
  1. Abstract: grammatical issues ('positive affect for models understand' should read 'positive effect on models' understanding of').
  2. Abstract: the introduction of silence detection as a common failure mode is not followed by any empirical comparison in the reported experiments.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments. They correctly identify gaps in validation and documentation that we will address. Below we respond point by point, indicating the revisions we will incorporate.

read point-by-point responses

  1. Referee: [—] Abstract: the assertion that the dataset 'mirrors real-life verbal exchanges, including overlaps and strategic silences' is unsupported because no quantitative validation (e.g., pause-duration histograms or overlap statistics) against any human-recorded Turkish corpus is reported; this directly undermines the claim that models trained on it learn genuine linguistic cues rather than artifacts of the generation process.

    Authors: We agree that the current abstract overstates the mirroring claim without supporting quantitative evidence. No public human-annotated Turkish turn-taking corpus exists for direct comparison, which is the core motivation for creating synthetic data. In revision we will (1) add internal statistics on the generated dataset (pause-duration histograms, overlap frequency, and silence length distributions) and (2) revise the abstract and introduction to state that the dataset is designed to incorporate these phenomena through LLM prompting rather than claiming empirical equivalence to human data. We will also add an explicit limitations paragraph on this point. revision: yes

  2. Referee: [—] Abstract / Results: all performance figures (BI-LSTM 0.839 accuracy, 0.910 AUC; Ensemble LR+RF) are obtained from models trained and tested on the same synthetic data with no held-out real Turkish test set and no comparison to the silence-detection baseline explicitly mentioned in the introduction; this circularity makes the generalization claim load-bearing and untested.

    Authors: We accept that all reported metrics are internal to the synthetic data and that a direct comparison to the silence-detection baseline is missing. We will add the baseline comparison (simple fixed-threshold silence detection) to the results section in the revision. However, no publicly available real Turkish dialogue corpus with turn-taking annotations exists for an external test set; creating one would require new data collection beyond the scope of this resource paper. We will clarify this limitation in the discussion and frame the current results as demonstrating learnability from the synthetic resource rather than proven generalization to real speech. revision: partial

  3. Referee: [—] Data generation description: the manuscript provides no details on prompting strategies, temperature settings, or post-processing used to elicit overlaps and strategic silences from Qwen models, nor any statistical fidelity checks, leaving the core premise that the synthetic data reproduces human Turkish turn-taking patterns unverified.

    Authors: We agree the generation details are insufficient. In the revised manuscript we will expand the data-generation section with: (a) the exact prompt templates used to elicit overlaps and strategic silences, (b) temperature and other sampling parameters (temperature = 0.8, top-p = 0.9), (c) post-processing rules applied to insert timing annotations, and (d) basic statistical fidelity checks comparing generated pause and overlap distributions against values reported in Turkish conversation analysis literature. revision: yes

standing simulated objections not resolved
  • External validation on a held-out real Turkish test set, because no suitable public annotated corpus currently exists.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper generates a synthetic Turkish dialogue dataset via Qwen LLMs and reports standard ML evaluation results (BI-LSTM accuracy 0.839, AUC 0.910) on train/test splits of that dataset. No equations, parameters, or first-principles derivations are presented that reduce by construction to the inputs. There are no self-citations invoked as load-bearing uniqueness theorems, no ansatzes smuggled via prior work, and no renaming of known results. The performance metrics are ordinary empirical outcomes from fitting models to the created data rather than predictions forced to equal the generation process. The interpretive claim that the dataset aids real linguistic-cue understanding is unsupported by external validation but does not constitute a circular reduction per the defined patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the unverified premise that Qwen LLM outputs replicate real Turkish conversational timing; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption Large language models can generate synthetic dialogues that sufficiently capture real Turkish turn-taking behaviors including overlaps and strategic silences.
    Invoked when creating the dataset but not supported by any validation metrics in the abstract.

pith-pipeline@v0.9.0 · 5473 in / 1294 out tokens · 53510 ms · 2026-05-10T13:27:06.513464+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Ekstedt and G

    E. Ekstedt and G. Skantze, ``TurnGPT: A Transformer-based Language Model for Predicting Turn-taking in Spoken Dialog,'' Findings of the Association for Computational Linguistics: EMNLP 2020, pp. 2981--2990, November 2020

    work page 2020

  2. [2]

    Ekstedt and G

    E. Ekstedt and G. Skantze, ``Projection of Turn Completion in Incremental Spoken Dialogue Systems,'' Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue, pp. 431--437, July 2021

    work page 2021

  3. [3]

    Roddy, G

    M. Roddy, G. Skantze, and N. Harte, ``Multimodal Continuous Turn-Taking Prediction Using Multiscale RNNs,'' Proceedings of the 20th ACM International Conference on Multimodal Interaction, pp. 186--190, October 2018

    work page 2018

  4. [4]

    Castillo-López, G

    G. Castillo-López, G. de Chalendar, and N. Semmar, ``A Survey of Recent Advances on Turn-taking Modeling in Spoken Dialogue Systems,'' Proceedings of the 15th International Workshop on Spoken Dialogue Systems Technology, pp. 254--271, May 2025

    work page 2025

  5. [5]

    A. Cano, G. Perez, L. Merino, and R. Gomez, ``Towards Improving Turn-Taking in Social Robots Using Visual-Only Voice Activity Detection in Multimodal Dialogue Systems,'' Social Robotics + AI: 17th International Conference, ICSR+AI 2025, Proceedings, Part II, pp. 207--221, September 2025

    work page 2025

  6. [6]

    Y. Lin, Y. Zheng, M. Zeng, and W. Shi, ``Predicting Turn-Taking and Backchannel in Human-Machine Conversations Using Linguistic, Acoustic, and Visual Signals,'' arXiv preprint arXiv:2505.12654, 2025

    work page arXiv 2025

  7. [7]

    Schuppler, M

    B. Schuppler, M. Hagmueller, J. A. Morales-Cordovilla, and H. Pessentheiner, ``GRASS: The Graz Corpus of Read and Spontaneous Speech,'' Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pp. 1465--1470, May 2014

    work page 2014

  8. [8]

    Eason, B

    G. Eason, B. Noble, and I. N. Sneddon, ``On certain integrals of Lipschitz-Hankel type involving products of Bessel functions,'' Phil. Trans. Roy. Soc. London, vol. A247, pp. 529--551, April 1955

    work page 1955

  9. [9]

    Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol

    J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford: Clarendon, 1892, pp.68--73

    work page

  10. [10]

    I. S. Jacobs and C. P. Bean, ``Fine particles, thin

    work page

  11. [11]

    Elissa, ``Title of paper if known,'' unpublished

    K. Elissa, ``Title of paper if known,'' unpublished

    work page

  12. [12]

    Nicole, ``Title of paper with only first word capitalized,'' J

    R. Nicole, ``Title of paper with only first word capitalized,'' J. Name Stand. Abbrev., in press

    work page

  13. [13]

    Yorozu, M

    Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, ``Electron spectroscopy studies on magneto-optical media and plastic substrate interface,'' IEEE Transl. J. Magn. Japan, vol. 2, pp. 740--741, August 1987 [Digests 9th Annual Conf. Magnetics Japan, p. 301, 1982]

    work page 1987

  14. [14]

    Young, The Technical Writer's Handbook

    M. Young, The Technical Writer's Handbook. Mill Valley, CA: University Science, 1989

    work page 1989

  15. [15]

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

    work page