Korean Two-Speaker Conversational Speech Datasets Compared

What a Korean full-duplex model needs

A full-duplex model listens and speaks at the same time, so it learns turn-taking from recordings where both people talk live and each voice sits on its own track. Moshi learned this from Fisher, about 2,000 hours of English phone calls with each side on its own channel (Défossez et al., 2024).

Timing does not carry over between languages. In the reference data of the J-Moshi paper, overlap runs 8.1 seconds per minute in Japanese and 3.3 seconds in English, though the figures come from different corpora (Ohashi et al., 2025). We did not find a published Korean figure. A model trained only on English has not heard how Korean speakers backchannel or interrupt while the other person talks. The fullduplex.ai page on training data covers the general requirements.

Korean full-duplex work so far

KRAFTON's Raon-Speech covers English and Korean. Its full-duplex extension, Raon-SpeechChat, trained on 119,000 hours of time-aligned real and synthetic dialogue, reported as one combined figure (Raon-Speech Technical Report). The released Raon-SpeechChat-9B checkpoint describes real-time conversation in English and is licensed CC BY-NC 4.0 (Hugging Face).

Synthetic data is the other route. A Korean full-duplex synthetic corpus of 89,273 conversations and 2,000.5 hours has a public 100-conversation preview (Hugging Face). The user side is generated with CosyVoice3 and the assistant side with Qwen3-TTS, and each conversation is labeled with normal turns, barge-ins, backchannels, and cut-offs.

The datasets side by side

Telephone and public corpora

The LDC's Korean Telephone Conversations has 100 calls between native speakers, each up to 30 minutes, with 15 to 18 minutes of each transcribed. Of these, 49 were released earlier as CALLFRIEND Korean and 51 were new. The audio is 2-channel 8 kHz mu-law, about 44 hours in total (LDC announcement, LDC2003S07). It is the closest Korean match to Fisher's format, but at telephone bandwidth, and the first 49 calls were published in 1996. Commercial use depends on your LDC membership terms.

KsponSpeech is a large public Korean corpus built from conversation. ETRI built it by "recording the dialogue of two people freely conversing on a variety of topics," 969 hours from about 2,000 speakers in a clean environment (Bang et al., 2020). It was designed for speech recognition. AI Hub distributes it as utterance-segmented files, cut at long pauses, in 16 kHz, 16-bit headerless PCM, with EUC-KR transcripts that mark fillers and disfluencies (AI Hub). Two limits matter for full-duplex work. The release gives you utterances, so check whether sessions and per-speaker timing can be rebuilt before you plan around it. And the AI Hub page states that only Korean nationals can apply for the data.

Open samples and synthetic data

MAGICDATA's Korean duplex set is 10 hours of two-speaker conversation recorded on mobile phones, with each speaker's audio on a separate track at 16 kHz (MagicHub). It is free to download under CC BY-NC-ND 4.0, and the page says commercial use is strictly prohibited. The ND term also limits sharing modified versions. Use it to evaluate a model or to get a first look at Korean overlap; the license rules out training a product on it.

The synthetic preview is useful for checking a pipeline that handles interruptions, since every event comes with a label and timestamps. The preview audio is 24 kHz mono, and the voices are TTS conditioned on the Zeroth-Korean read-speech corpus. Its overlap timing comes from a generator, so pair it with recorded conversation.

Commercial datasets

Nexdata's Korean set is 16 kHz smartphone audio from about 200 speakers. The page title says 200 hours and the product name says 214, and it lists the files as mono without saying how speakers map to files (Nexdata). Ask for the layout and the hours in the quote.

SpeechData.ai lists 500 hours of spontaneous two-person Korean with one channel per speaker, 50 speakers, time-aligned human-verified transcripts, and signed commercial-use consent from every speaker, at $95 per hour (Hugging Face card). The card does not give a sample rate. Fifty speakers across 500 hours means many hours per voice, so check speaker variety against what your model needs.

Basis Conversations 1500 has 9.5 hours of Korean in 28 conversations with 35 speakers. Each speaker has a separate 48 kHz track on a shared clock, but these are group calls with up to four people talking at once (Hugging Face). Access is gated, the license is Basis's own, and the card warns against training on its bundled transcripts.

otoSpeech Korean is free conversation between two people on a given topic, recorded at 44.1 kHz with one speaker on each channel and ASR transcripts for both speakers (oto). The transcripts have no manual review, and the dataset has no turn-taking labels. A sample on the topic "Talk freely about AI" lets you switch off either speaker and hear one channel alone.

What to check in a Korean sample

• Separate at the source: Play a stretch where both people talk and listen to each channel alone. You should hear one voice per channel, with the other faint or absent.
• Real bandwidth: A 44.1 or 48 kHz file with almost no energy above 8 kHz was probably upsampled from a lower rate.
• Whole sessions: Utterance files lose the gaps and overlaps between speakers. Ask for full sessions with timestamps per channel.
• Transcripts: Find out whether transcripts are ASR output or human-verified, and whether backchannels and overlapped words are kept.
• License and access: Check for non-commercial or no-derivatives terms, nationality limits on access, and written consent for AI training.

The guide to buying full-duplex speech data includes a short script that measures channel separation, overlap, and bandwidth on a sample.

Where oto fits

• Korean two-speaker conversation. otoSpeech Korean is free conversation on a given topic, 44.1 kHz stereo with one speaker per channel and ASR transcripts for both speakers. Listen on oto.earth/datasets/korean.
• Licensing. Korean data is licensed on request. oto's licensed datasets run from $10 to $100 per hour of audio, and the license and labels such as turn-taking annotation add to the price.
• Collection to specification. Other topics, scenarios, speakers, manual transcripts, and turn-taking labels, quoted per project. Contact agent@oto.earth.
• Same format in other languages. English and Japanese conversation with one speaker per channel, listed on oto.earth/datasets.

Frequently asked questions

Is there an open Korean full-duplex conversation dataset?

MAGICDATA's 10-hour Korean duplex set at 16 kHz is free to download but non-commercial; the larger per-speaker sets we found are commercial.

Can I use KsponSpeech to train a full-duplex model?

It was recorded as two-person conversation but is released as utterance files for ASR, and AI Hub accepts applications from Korean nationals only.

Where can I buy Korean two-speaker conversational speech data?

From vendors such as oto, SpeechData.ai, and Nexdata; ask each for a sample, the channel layout, and the license text.

How much does Korean conversational speech data cost?

SpeechData.ai lists $95 per hour, and oto's licensed datasets run from $10 to $100 per hour depending on the license and labels.

What data did KRAFTON use for Raon-SpeechChat?

The technical report gives 119,000 hours of time-aligned real and synthetic dialogue as one combined figure, without a Korean breakdown in its abstract.

What sample rate is otoSpeech Korean?

It is 44.1 kHz stereo, with one speaker on each channel and ASR transcripts for both speakers.

For more datasets, benchmarks, and models, see the fullduplex.ai catalog.

Related guides

• Japanese Conversational Speech Datasets for Full-Duplex Models
• How to Buy Full-Duplex Conversational Speech Data
• Real vs. Synthetic Dialogue for Speech-to-Speech Training