When Li Mu and Alex Smola's team developed the WildASR benchmark, they specifically selected MagicHub's natural full-duplex conversational speech data as a core data source. This not only represents recognition of our data quality by leading researchers, but also reinforces a fundamental insight: Real-world speech interaction must be learned from real human conversations, including hesitations, interruptions, and disfluencies.
That's why we've decided to continue making our datasets open source.
This time, we're introducing the Korean Full-Duplex Natural Conversation Dataset: MAGICDATA-DUPLEX-KOREAN.
Why Korean?
Korean has distinctive characteristics in natural conversation. Honorific levels shift depending on speakers and social contexts, rich sentence endings convey emotions and attitudes, and frequent backchannels, turn-taking, and overlapping speech are integral to everyday communication.
For full-duplex speech models, the challenge goes beyond understanding Korean. Models must also learn when to respond, when to listen, and how to communicate naturally across different conversational contexts.
However, read speech and turn-based dialogues still account for a significant portion of existing speech datasets. High-quality full-duplex data that preserves the complete dynamics of real human interaction remains a valuable and relatively scarce resource.
What Makes Our Dataset Different?
Real Conversations, Not Synthetic Speech.
All recordings come from natural conversations between real people, preserving hesitations, backchannels, interruptions, overlapping speech, and immediate feedback. These are essential elements for models to learn authentic human interaction.
Native Dual-Track Recording, Independently Captured.
Each speaker's audio is recorded on a separate track, preserving the complete interaction between participants. No post-processing speech separation is required, making the data particularly suitable for full-duplex model training.
Carefully Selected, Highly Interactive Samples.
We select conversation segments with rich interaction patterns, focusing on immediate responses, turn transitions, natural interruptions, and other key conversational behaviors.
Authentic Korean Conversational Context.
The dataset captures honorific expressions, variations in tone, sentence-ending features, and natural turn-taking patterns, providing authentic examples to help models understand Korean conversational conventions.
Compliant Data Collection and Full Copyright Ownership.
All data has been collected in compliance with applicable requirements, with full rights to the data secured.
Who Is This Dataset For?
This dataset is suitable for research involving Korean full-duplex speech models, real-time Voice Agents, voice assistants, AI companion applications, interruption detection, turn-taking prediction, overlapping speech processing, and other related fields.
Usage Notice: This dataset is released for non-commercial use only. It may be used for academic research, education, and non-commercial model training and evaluation. Commercial use is strictly prohibited. Please refer to the license terms on the dataset's official page for detailed usage requirements.
Why Magichub.com?
While there are many open-source data communities, dataset quality varies significantly. Many datasets appear impressive in scale, but once downloaded, they require extensive additional processing— inconsistent formats, non-standard annotations, misaligned audio and text, uncleaned noise, and missing metadata. Very few are truly ready for direct use in model training.
What sets Magichub.com apart is that all our datasets are designed to be ready to use out of the box. You don't need to spend additional time and resources on data cleaning, alignment, format conversion, or annotation correction. Every data sample undergoes rigorous quality inspection and standardization before release, allowing it to be used directly for training, evaluation, or product development.
For AI teams working under tight deadlines, this means saving weeks of data preprocessing and gaining faster iteration cycles and a more reliable starting point for experimentation.
MagicHub is becoming the foundation of real-world data for the Voice AI era.
