Paper deep dive
Syn-TurnTurk: A Synthetic Dataset for Turn-Taking Prediction in Turkish Dialogues
Ahmet Tuğrul Bayrak, Mustafa Sertaç Türkel, Fatma Nur Korkmaz
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/18/2026, 1:29:10 AM
Summary
Syn-TurnTurk is a synthetic Turkish dialogue dataset designed to improve turn-taking prediction in voice-based chatbots. Generated using various Qwen LLMs, the dataset addresses the lack of high-quality Turkish resources by capturing natural speech patterns, including overlaps and strategic silences. Experimental results using BI-LSTM and Ensemble (LR+RF) models demonstrate high accuracy (0.839) and AUC (0.910), confirming the dataset's utility for training models to recognize linguistic cues in Turkish conversations.
Entities (5)
Relation Signals (3)
Syn-TurnTurk → targets → Turkish
confidence 100% · Syn-TurnTurk, a synthetic Turkish dialogue dataset
BI-LSTM → evaluatedon → Syn-TurnTurk
confidence 95% · We evaluated the dataset using several traditional and deep learning architectures... BI-LSTM... achieve high accuracy
Syn-TurnTurk → generatedby → Qwen
confidence 95% · Syn-TurnTurk, a synthetic Turkish dialogue dataset generated using various Qwen Large Language Models
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2604.13620v1
- Canonical: https://arxiv.org/abs/2604.13620v1
Trouble viewing inline? Open PDF directly →
Full Text
15,036 characters extracted from source content.
Expand or collapse full text
Syn-TurnTurk: A Synthetic Dataset for Turn-Taking Prediction in Turkish Dialogues Ahmet Tu ̆ grul Bayrak Data Science and Innovation Ata Technology Platforms ̇ Istanbul, Turkey tugrul.bayrak@atptech.com Mustafa Sertac ̧ T ̈ urkel Data Science and Innovation Ata Technology Platforms ̇ Istanbul, Turkey sertac.turkel@atptech.com Fatma Nur Korkmaz Data Science and Innovation Ata Technology Platforms ̇ Istanbul, Turkey fatmanur.korkmaz@atptech.com This is the author’s version of a paper accepted for publication in IEEE ICASI 2026. 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 inter- rupt 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. Index Terms—turn-taking prediction, synthetic dataset, pre- dictive modeling, turngpt, qwen I. INTRODUCTION The growth of generative AI has made chatbots a common tool in many industries. While creating a basic system is now straightforward, largely due to methods like Retrieval- Augmented Generation (RAG), developing a bot that interacts like a human remains a significant challenge. This difficulty is especially clear in voice-based systems, where the timing of a conversation is essential for a natural experience. Most voice-enabled chatbots function by monitoring for a specific duration of silence. Once the user stops speaking for a few seconds, the system assumes the turn has ended and begins its response. However, this method is often unreliable because speech patterns vary between individuals. People frequently pause in the middle of a sentence or between words. If a bot interprets these pauses as a finished turn, it will interrupt the user. This mistake disturbs the conversational flow and makes the interaction feel mechanical rather than human. A fundamental requirement for realistic spoken dialogue systems is the ability to manage turn-taking with human-like speed and accuracy. Recent surveys highlight that a major challenge in this field is the lack of established benchmarks to track progress and compare models on a standard ground [4]. Understanding the underlying structure of speech is es- sential for modeling these interactions. The GRASS corpus [7] established a framework for this by defining turn-taking in terms of layers such as Inter-Pausal Units (IPU) and Potential Completion Points (PCOMP). These linguistic markers served as a core motivation for TurnGPT [1], a study that proved Transformer-based models could use dialogue context not only to detect turn changes but also to anticipate them. Building on this, ”projection” mechanisms was introduced [2] designed to predict future completion points, which successfully reduced system latency and made interactions more immediate. As the field moves toward more complex systems, models must handle signals that occur at different temporal speeds. Roddy et al. [3] addressed this by proposing a Multiscale RNN architecture that processes linguistic and acoustic features at separate rates. More recently, the scope of turn-taking has ex- panded to include multimodal signals. For instance, the M- F2F dataset [6] combines linguistic, acoustic, and visual data to improve the prediction of both turn-taking and backchannel actions. Visual information is particularly useful when audio data is compromised or unavailable. In such scenarios, Cano et al. [5] demonstrated that social robots can rely on Visual Voice Activity Detection (VVAD) to identify speech boundaries and manage conversational flow using only visual cues. The difficulty of predicting turns is even greater for lan- guages that are not as widely supported in global datasets. Turkish, in particular, suffers from a lack of high-quality, labeled conversation data needed to train accurate turn-taking models. Due to its unique sentence structure and suffix-based grammar, existing models trained on English often fail to capture the nuances of Turkish dialogue. To address this gap, this paper presents a synthetic Turkish dialogue dataset. The goal is to train models that can recognize turn boundaries by analyzing the linguistic structure of Turkish conversations. By learning from these computer-generated examples, the models can perceive when a user has completed their thought based on the specific words and grammatical markers used. This approach moves away from a reliance on simple silence and toward a more intelligent, language-specific understanding of conversational flow.© 2026 IEEE. Personal use of this material is permitted. arXiv:2604.13620v1 [cs.CL] 15 Apr 2026 I. DATASET GENERATION The construction of this dataset involved the utilization of five distinct Qwen models via API calls: qwen3-max-2026- 01-23, qwen3.5-35b-a3b, qwen3.5-plus-2026-02-15, qwen3.5- 397b-a17b, and qwen3.5-flash-2026-02-23. Each model was tasked with generating natural, two-person dialogues based on specific constraints to ensure the dataset’s diversity. To prevent semantic repetition and ensure a broad coverage of conversational contexts, a pool of 79 unique topics was established. For each API request, a topic was selected at random, serving as the thematic foundation for the interaction. Furthermore, the models were specifically instructed to incor- porate human-centric speech characteristics, such as overlaps, strategic silences, and everyday interjections, to mirror real-life verbal exchanges as closely as possible. The variability of the generated outputs was controlled by adjusting the temperature parameter. Although different values were used to increase diversity, most generations were performed at a temperature of 0.7 to balance coherence and conversational spontaneity. The resulting raw data (Syn-TurnTurk) was subsequently formatted and hosted on Hugging Face 1 for further analysis. I. DATASET CHARACTERISTICS AND STRUCTURAL ANALYSIS A structural analysis was conducted to quantify the dataset’s characteristics and ensure its suitability for training predictive models. The final corpus consists of 1,625 dialogues, with a total of 12,560 individual speaker changes. This high volume of turn-taking suggests that the interactions are sufficiently dynamic, moving beyond simple prompt-response pairs into more complex, multi-turn exchanges. The dataset’s diversity is reflected in both its thematic breadth and its parametric variety. By distributing the dialogues across 79 distinct topics, the risk of linguistic over-fitting was minimized. The distribution of dialogues across the five Qwen models is detailed in Table I. TABLE I DISTRIBUTION OF DIALOGUES BY MODEL Model NameCountPercentage (%) qwen3-max-2026-01-2367541.5% qwen3.5-35b-a3b28317.4% qwen3.5-flash-2026-02-2327016.6% qwen3.5-397b-a17b22814.0% qwen3.5-plus-2026-02-1516910.4% To evaluate the timing of the conversations, Floor Trans- fer Offset (FTO) and general interaction flow metrics were measured to capture the temporal nuances of human-like exchanges. These values, summarised in Table I, represent the core temporal characteristics and the dynamic nature of the generated dialogues. As illustrated in Fig. 1, the difference between the mean and median FTO suggests an asymmet- ric distribution, with a significant number of longer silence gaps alongside frequent rapid transitions, which are critical 1 https://huggingface.co/datasets/tugrulbayrak/Syn-TurnTurk for modelling natural speech patterns accurately. Specifically, the dataset contains 5,305 documented instances of overlaps, totalling 2,213.50 seconds of concurrent speech, reflecting a high level of interactivity within the synthetic interactions. These metrics provide a robust baseline for the subsequent evaluation of 30 different combinations of models, specifically testing their ability to distinguish between intentional pauses and actual completed turns. TABLE I SUMMARY OF TIMING AND INTERACTION METRICS MetricValue Mean FTO0.286s Median FTO0.743s Max Overlap (Negative FTO)-2.500s Max Silence Gap (Positive FTO)0.880s Total Number of Overlaps5,305 Avg. Overlaps per Dialogue3.26 Avg. Silence per Dialogue3.58s Fig. 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 turn transition, the final one-third of the text sequence was labeled as 1, while two distinct segments from the remaining portion were randomly selected and labeled as 0. The final dataset consists of 12,560 positive and 25,120 negative samples. To ensure comparability across model subsets, the training data was downsampled to match the smallest subset size. The re- sulting training set contained 1,306 positive and 2,696 negative samples. For text representation, the intfloat/multilingual-e5- large embedding model was applied, as it is highly effective at capturing the semantic and structural nuances of the Turkish TABLE I OVERALL PERFORMANCE COMPARISON: ABSOLUTE BEST MODEL PER METRIC HIGHLIGHTED Data SubsetModelPrecisionRecallF1-ScoreAccuracyAUC Full DatasetLogistic Regression0.6970.8450.7640.8160.898 Decision Tree0.5970.6000.5980.7210.687 Random Forest0.8080.5640.6640.8020.890 Ensemble (LR+RF)0.7540.7940.7730.8360.907 LSTM0.7650.7610.7630.8380.904 BI-LSTM0.7670.7780.7720.8380.905 qwen3.5-397b-a17bLogistic Regression0.6610.8130.7290.7940.882 Decision Tree0.5710.5960.5840.7090.634 Random Forest0.8070.4560.5820.7770.867 Ensemble (LR+RF)0.7370.7230.7300.8180.886 LSTM0.7510.7190.7350.8230.892 BI-LSTM0.7520.6890.7190.8160.889 qwen3.5-plus-2026-02-15Logistic Regression0.6770.8210.7420.8060.887 Decision Tree0.5520.5970.5740.6980.631 Random Forest0.8140.4540.5830.7790.868 Ensemble (LR+RF)0.7490.7260.7370.8240.886 LSTM0.7180.7510.7340.8150.888 BI-LSTM0.7420.6720.7050.8090.887 qwen3-max-2026-01-23Logistic Regression0.6810.8370.7510.8130.897 Decision Tree0.5850.6840.6310.7300.686 Random Forest0.8110.5440.6510.8040.893 Ensemble (LR+RF)0.7470.7850.7660.8380.910 LSTM0.7860.7100.7460.8370.906 BI-LSTM0.7760.7340.7540.8390.907 qwen3.5-35b-a3bLogistic Regression0.6870.8280.7510.8140.893 Decision Tree0.5700.6380.6020.7150.648 Random Forest0.8050.4700.5940.7820.874 Ensemble (LR+RF)0.7330.7330.7330.8190.893 LSTM0.7610.7430.7520.8340.894 BI-LSTM0.7610.7430.7520.8340.895 language. Furthermore, the performance of each model was evaluated using a 5-fold cross-validation approach. The spe- cific hyperparameters and configurations for each model are detailed below: • DT: crit = Gini, split = Best, minsplit = 2 • LR: pen = L2, sol = lbf gs, iter = 1000, C = 1.0 • RF: est = 100, crit = Gini, depth = N one, boot = T • LSTM: hid = 384, opt = Adam, lr = 0.001 • BI-LSTM: hid = 384, opt = Adam, lr = 0.001 • Ensemble (LR+RF): vote = Sof t, weight = Equal V. CONCLUSION Turn-taking remains a significant challenge in conversa- tional AI. While many chatbots have been developed, predict- ing exactly when a user has finished speaking is still difficult. In this study, we created a natural Turkish dialogue dataset using various Qwen LLM models and temperature settings to address the lack of available resources for the Turkish language. We then evaluated this dataset using several machine learning and deep learning models. The experimental results, shown in Table I, demonstrate that the dataset provides a strong foundation for training turn-taking models. The BI- LSTM and Ensemble (LR+RF) models delivered the most balanced performance across all metrics. Specifically, the BI- LSTM model achieved the highest accuracy of 0.839, while the Ensemble model reached a peak AUC of 0.910. Notably, the slightly lower performance observed in the qwen3.5-397b- a17b subset suggests that the increased linguistic complexity and more natural conversational flow of advanced models make turn-taking points harder to predict for classifiers com- pared to simpler outputs. Furthermore, the high performance of the LSTM-based architectures confirms that understanding the linguistic flow is essential for managing transitions in Turkish speech. While simpler models can identify potential turn-ending points, the advanced models provide the stability needed for natural interaction. These results demonstrate that the synthetic dataset might be useful for real-world dialogue predictions. REFERENCES [1] 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, p. 2981– 2990, November 2020. [2] E. Ekstedt and G. Skantze, “Projection of Turn Completion in Incremen- tal Spoken Dialogue Systems,” Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue, p. 431–437, July 2021. [3] 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, p. 186–190, October 2018. [4] G. Castillo-L ́ opez, G. de Chalendar, and N. Semmar, “A Survey of Re- cent Advances on Turn-taking Modeling in Spoken Dialogue Systems,” Proceedings of the 15th International Workshop on Spoken Dialogue Systems Technology, p. 254–271, May 2025. [5] A. Cano, G. Perez, L. Merino, and R. Gomez, “Towards Improving Turn-Taking in Social Robots Using Visual-Only Voice Activity De- tection in Multimodal Dialogue Systems,” Social Robotics + AI: 17th International Conference, ICSR+AI 2025, Proceedings, Part I, p. 207– 221, September 2025. [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. [7] B. Schuppler, M. Hagmueller, J. A. Morales-Cordovilla, and H. Pessen- theiner, “GRASS: The Graz Corpus of Read and Spontaneous Speech,” Proceedings of the Ninth International Conference on Language Re- sources and Evaluation (LREC’14), p. 1465–1470, May 2014.