Easy-Turn-Trainset
1,145-hour Mandarin training dataset for full-duplex dialogue turn-taking detection, combining real and synthetic data
Duration
1145 hours
Languages
1
Sample Rate
None
Published
2025-09
Description
1Total duration of approximately 1,145 hours, containing four dialogue states: complete state (580 hours, 423k samples), incomplete state (532 hours, 712k samples), backchannel state (10 hours, 41k samples), and wait state (23 hours, 40k samples)
2Includes real data (sourced from MagicData-RAMC Mandarin conversational corpus) and synthetic data (text generated by DeepSeek V3 / Qwen2.5-72B + CosyVoice 2 speech synthesis)
3Synthetic data verified by Paraformer to achieve 0% WER quality standard
4Designed for training turn-taking detection models in full-duplex dialogue systems to determine when users finish speaking
5Apache-2.0 license
Language Details
| Language | Duration |
|---|---|
| Mandarin Chinese | 1145 hours |
Publisher