Japanese Conversational Speech Datasets for Full-Duplex Models

日本語で読む

Why Japanese needs its own two-channel data

Moshi learned to listen and speak at once from Fisher, about 2,000 hours of English phone calls with each side on its own channel (Défossez et al., 2024). That timing does not transfer cleanly to Japanese. In the reference data of the J-Moshi paper, overlap runs 8.1 seconds per minute in Japanese and 3.3 seconds in English (Ohashi et al., 2025). The two numbers come from different corpora, so treat the gap as indicative.

The J-Moshi authors point to earlier studies that found more overlap and backchannels in Japanese conversation. A model can only learn which of these are interruptions and which are not if both voices are clean on separate tracks. The fullduplex.ai page on training data covers the general requirements.

What Japanese full-duplex models trained on

J-Moshi, from Nagoya University, adapted Moshi to Japanese in three stages (Ohashi et al., 2025):

• Pre-training: J-CHAT, about 69,000 hours of podcast and YouTube audio. J-CHAT is mono, so the team ran speaker diarization and put one speaker on Moshi's channel and the rest on the user's channel.
• Fine-tuning on 344 hours of real two-channel dialogue: Travel Agency Task Dialogue (Tabidachi) 330 dialogues and 115 h, an in-house Casual Dialogue Corpus 500 dialogues and 148 h, an in-house Consultation Dialogue Corpus 100 dialogues and 53 h, Japanese CALLHOME 120 dialogues and 16 h, and CSJ 58 dialogues and 12 h.
• Synthetic extension: J-Moshi-ext adds 602 hours of two-channel speech generated by multi-stream TTS.

The two in-house corpora account for 201 of the 344 hours, and they are not released. Nobody can rebuild J-Moshi's fine-tuning mix from public sources.

LLM-jp-Moshi-v1, released by the National Institute of Informatics (NII) on February 25, 2026, is under Apache 2.0 and was announced as the first commercially usable Japanese full-duplex model (NII). It pre-trained on J-CHAT and fine-tuned on LLM-jp-Zoom1, about 1,000 hours of Zoom conversation that NII said would be released soon. As of November 9, 2026, we could not find LLM-jp-Zoom1 on Hugging Face. The fullduplex.ai profile of the release covers the background.

Academic corpora

CSJ (Corpus of Spontaneous Japanese) is a monologue corpus with a small dialogue part. Task-oriented dialogue, free dialogue, and interviews add up to about 12 of its 658.8 hours, recorded with head-worn close-talking microphones and down-sampled to 16 kHz (Maekawa). The commercial tier is described as "for the purpose of commercialising the corpus statistics" (CSJ fees), and the page does not mention model training. Ask NINJAL before you train on it.

CEJC (Corpus of Everyday Japanese Conversation) was recorded in daily life by volunteers. Each main speaker wore an IC recorder with a built-in microphone, and the released files are 16 kHz mono per recorder plus a central stereo file (CEJC audio). Built-in microphones close together will pick up other speakers, so check bleed. Commercial licenses run from ¥500,000 for two years and up to 10 users, excluding tax, and the contract page does not mention AI training (CEJC contract).

BTSJ collects first meetings, friends' conversations, and thesis supervision sessions of about 20 minutes each. Of the 514 conversations, 384 include audio (NINJAL). Use is limited to education and research.

CALLHOME Japanese came out in a Second Edition on January 15, 2026, with 16-bit FLAC audio converted from the original 8 kHz files and revised transcripts (LDC2026S02). Commercial use depends on your LDC membership terms.

Tabidachi is the largest obtainable Japanese role-play set, recorded over Zoom with 5 operators who had travel agency experience and 55 customers, 20 of them children. The IDR page lists the release as MP4 video and JSON annotations, so confirm the audio layout when you apply (NII IDR). Private companies are reviewed case by case.

ATR 48k is a 1994 to 1997 role-play corpus. Speakers improvised from a conversation plot, without reading, and were not face to face (ATR spec). It is one of the few with a commercial license (¥4,000,000 excluding tax), but it ships as mono files per utterance with 500 ms to 1 s of padding. Ask whether the session timeline can be rebuilt, because overlap timing depends on it.

TTS-oriented and web-scale sets

STUDIES is the clearest open set of acted Japanese dialogue, but its speakers were recorded "on separate days for each speaker" as a COVID-19 measure (Saito et al., 2022). It has no real overlap. Its call-center extension records the operator side only.

J-CHAT totals 76,036 hours in its paper (Nakata et al., 2024). Its card limits use to information analysis under Article 30-4 of Japan's Copyright Act and does not allow commercial use (Hugging Face). ReazonSpeech, more than 35,000 hours from TV, carries the same Article 30-4 limit (Reazon). DuplexChat publishes 132,723 hours of Japanese clip URLs and timestamps and no audio, so you download and separate the episodes yourself.

Commercial datasets

Vendor specs are self-reported, so check them in a sample. Nexdata's 579-hour set is the largest 48 kHz free-conversation listing we found, though its page lists the format as mono, so ask how speakers map to files. SpeechData.ai publishes a price and states signed commercial-use consent from every speaker, but no sample rate. The MagicData set is free to download and does not allow commercial use (MagicHub). Basis Conversations 1500 has 51 Japanese conversations with 62 speakers, and its card warns against training on the bundled transcripts (Hugging Face).

Role-play, free conversation, or scripted

The conversation style shows up in the timing. Abe and colleagues counted overlaps per 20 seconds: 5.54 in LLM-jp-Zoom1 free conversation, 2.46 in CALLHOME, and 1.16 in Tabidachi role-play (Abe et al., ASJ 2026). In Tabidachi one side produced 5.06 speech units per 20 seconds and the other 2.12, so the operator did most of the talking. When you evaluate role-play data, check that the scenario keeps both people talking.

oto measured its own sample clips the same way. Speech time with overlap was 19% in free conversation, 10% in role-play, and 0% in scripted dialogue (oto). These are short clips picked to show the data, not full-set statistics. Scripted dialogue will not teach turn detection under overlap, but its clean turns and exact scripts suit TTS.

Sample rate matters most for TTS. Moshi's audio tokenizer, Mimi, runs at 24 kHz (Ohashi et al., 2025), so 48 kHz audio is more than a Moshi-class model uses, while 8 kHz telephone audio is less.

Where oto fits

• Three Japanese styles, one format. otoSpeech Japanese, Role-Play, and Scripted record two people talking in the same session, each on their own 48 kHz channel. Samples are on oto.earth/datasets/japanese (日本語).
• Measured on the samples. The otoSpeech Japanese research page reports overlap and short responses per style.
• Licensing. Japanese 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 scenarios, speakers, and turn-taking labels, quoted per project. Contact agent@oto.earth.

Frequently asked questions

What Japanese datasets did J-Moshi train on?

J-CHAT for pre-training, then 344 hours of two-channel dialogue: Tabidachi, CALLHOME Japanese, CSJ, and two in-house Nagoya corpora that make up 201 of those hours.

Is there a free Japanese full-duplex dataset I can use commercially?

We found none with real per-speaker recordings: the MagicData set is non-commercial, and LLM-jp-Zoom1 was not public as of November 9, 2026.

Does CSJ have two-speaker dialogue?

Yes, but only about 12 of its 658.8 hours are dialogue or interviews, at 16 kHz, and its commercial tier may not cover model training.

Can I train a commercial model on J-CHAT or ReazonSpeech?

Both limit use to information analysis under Article 30-4 of Japan's Copyright Act, and J-CHAT's card says commercial use is not allowed.

Is role-play data good for full-duplex training?

It adds task scenarios, but Tabidachi role-play had far fewer overlaps than free conversation, so check that both speakers talk actively in the samples.

Where can I buy Japanese two-channel conversation data?

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

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

Related guides

• Korean Two-Speaker Conversational Speech Datasets Compared
• Free Conversation, Role-Play, or Scripted: What Each Teaches a Speech Model
• How Much Full-Duplex Data Do You Need? What Published Systems Used