otoSpeech Korean: full-duplex conversation in Korean

Listen to a sample

This clip comes from a conversation on the topic “Talk freely about AI.” Switch off either speaker in the player to hear one channel alone.

• Talk freely about AI

Why Korean needs its own data

A voice system picks up its timing from the conversations it trains on, and full-duplex models started with English. Moshi learned to listen and speak at once from Fisher, 2,000 hours of phone calls with each side on its own channel.[1] Research on Japanese shows why that does not transfer: in the reference data in the J-Moshi paper, overlap runs 8.1 seconds per minute in Japanese and 3.3 in English, though those come from different corpora.[2] A model trained on English alone has not heard how Korean speakers respond while the other person talks.

Korean work is under way. KRAFTON’s Raon-Speech covers English and Korean, and its full-duplex extension trained on 119,000 hours of real and synthetic dialogue, reported as one combined figure.[3] For real recordings, what we found is mostly telephone or commercial data: 100 phone conversations from the Linguistic Data Consortium, and a Nexdata set of about 200 hours at 16 kHz.[4][5] We did not find an open Korean conversation corpus with separate channels at this sample rate, though that does not prove none exists.

What is included

• Stereo audio at 44.1 kHz, with one speaker on each channel
• ASR transcripts for both speakers

The transcripts have no manual review, and the dataset has no turn-taking labels. If your work needs them, we can add them to your specification.

fullduplex.ai, a site oto runs, lists this dataset alongside other full-duplex datasets.