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

2026-01-23 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, medium

Summary

Syn-TurnTurk is a new synthetic Turkish dialogue dataset designed to improve turn-taking prediction in voice-based chatbots, addressing the critical lack of high-quality, labeled conversational data for Turkish. Generated using five Qwen Large Language Models (LLMs) including qwen3-max-2026-01-23 and qwen3.5-397b-a17b, the dataset comprises 1,625 dialogues across 79 unique topics, featuring 12,560 speaker changes and 5,305 instances of speech overlaps. The generation process incorporated human-centric speech characteristics like strategic silences and interjections, with a temperature of 0.7 for most outputs. Evaluation using traditional and deep learning models, including Decision Tree, Random Forest, Logistic Regression, and BI-LSTM, showed that BI-LSTM achieved the highest accuracy of 0.839, while an Ensemble (LR+RF) model reached a peak AUC of 0.910, demonstrating the dataset's effectiveness for training models to understand linguistic cues.

Key takeaway

For research scientists developing conversational AI for less-resourced languages, this work demonstrates that synthetic data generation with LLMs can effectively create high-quality datasets for complex tasks like turn-taking prediction. You should consider using models like BI-LSTM or Ensemble methods, which showed superior performance on the Syn-TurnTurk dataset, to build more natural and responsive voice-based systems, especially for languages with unique grammatical structures like Turkish.

Key insights

Synthetic datasets can effectively address data scarcity for turn-taking prediction in under-resourced languages like Turkish.

Principles

• Dialogue timing is crucial for natural chatbot interaction.
• Linguistic flow is key for managing speech transitions.
• Synthetic data can mirror real-life verbal exchanges.

Method

Generate diverse two-person dialogues using multiple LLMs with varied topics and controlled temperature, incorporating human speech characteristics like overlaps and silences, then label turn transitions for classification.

In practice

• Use Qwen LLMs for synthetic dialogue generation.
• Incorporate 79+ topics to ensure dataset diversity.
• Apply BI-LSTM for high accuracy in turn prediction.

Topics

• Turn-Taking Prediction
• Synthetic Dataset
• Turkish Dialogues
• Qwen LLMs
• BI-LSTM

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Syn-TurnTurk: A Synthetic Dataset for Turn-Taking Prediction in Turkish Dialogues — Synthetic datasets can effectively train models for complex linguistic tasks like turn-taking prediction in under-resourced languages.
• TurnGuide: Enhancing Meaningful Full Duplex Spoken Interactions via Dynamic Turn-Level Text-Speech Interleaving — TurnGuide enhances full-duplex SLMs by planning turn-level text guidance before speech, improving semantic coherence and natural flow.
• Beyond Cascaded Pipelines: Building a Native Spoken Language Model Prototype — Native Spoken Language Models (SLMs) overcome cascaded pipeline limitations by directly processing audio, preserving nuance, reducing errors, and lowering latency.
• 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.
• Issue #120 - Turn-based voice AI agents — Voice agents fall into turn-based (sequential) or streaming (overlapping) architectures, each with distinct trade-offs.
• EMSDialog: Synthetic Multi-person Emergency Medical Service Dialogue Generation from Electronic Patient Care Reports via Multi-LLM Agents — Synthetic multi-speaker EMS dialogues generated from ePCRs enhance conversational diagnosis prediction accuracy and stability.

Open in AIssential →

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