Sound familiar?Need real back-and-forth conversation, not read speech?
Conversational Speech, Speaker-Separated
Unscripted Korean multi-party conversation. Overlaps, laughter and backchannels are all there, with one channel per speaker.
Use cases
Problems this data solves
Train full-duplex voice models on real turn-taking
Diarize even overlapping speech
Expressive voices, not read-aloud speech
Scale · specs
- Total volume
- [Hours] · [Clips]
- Speakers
- 2–6 people, Korean
- Genre
- Talk show · variety · radio · [other genres]
- Audio
- Per-speaker WAV, 48 kHz, one channel per speaker
- Conversation
- Overlap ratio · laughter and backchannel events marked
Labels
- Included
- Transcript, speaker timeline, overlap segments, laughter and backchannel events
- On request
- Emotion, other custom labels
Request a sample
Get a sample of this collection
The button takes you to the sample request form.
Delivery
How you receive it
File formatPer-speaker WAV + JSON transcript
Bundle unitPer conversation session
Delivery methodDownload link · cloud bucket (manifest included)
Rights & consent
Rights & consent
Direct contract with the rights holder
Contracted directly with the producer who owns the originals
Performer consent scope
Within the scope of performers’ consent to use their voice for AI training
Face handling options
Blur non-consenting faces or fully anonymize
Non-exclusive · fixed term
Non-exclusive use within the contract term
No-replication clause
No replicating performers’ faces or voices
Need something slightly different?
Change the speaker count, genre or labels and request a custom build.