otoSpeech Japanese: free conversation, role-play, and scripted dialogue

日本語で読む

Listen to samples

Press play and the player lists every topic or scenario in that set. Scripted clips show the script as they play.

• Free conversation: five topics
• Role-play: eight scenarios
• Scripted: four scenes

Why Japanese needs its own data

Full-duplex research started in English. Moshi, the first real-time full-duplex model, learned to listen and speak at once from Fisher: 2,000 hours of phone calls with each side on its own channel.[1] When Ohashi and colleagues built J-Moshi in 2025, they wrote that no such system existed for other languages and that findings outside English were lacking.[2] NII released LLM-jp-Moshi-v1 for commercial use in 2026, so this is moving.[3]

The data gap shows in J-Moshi’s training. It fine-tuned on 344 hours of stereo dialogue, and 201 of those hours came from the lab’s in-house corpora.[2] Speech data more broadly leans the same way: of 38 public speech datasets in a 2025 review, 15 were English-dominated and 9 were Mandarin.[4]

Timing learned from English does not carry over cleanly. In the reference data in the J-Moshi paper, overlap runs 8.1 seconds per minute in Japanese and 3.3 seconds in English.[2] Those figures come from different corpora, so the gap is indicative, but a model that learned English timing has heard far less of it.

Natural speech and specialist content pull against each other

On a familiar topic, people speak naturally and clearly, but specialist content rarely comes up. A script can put that content in, and reading it costs some natural delivery. We record three styles so you can pick your point on that line.

• For natural speech, use free conversation.
• For a specific domain, or for TTS, use scripted dialogue.
• For both natural speech and specialist content, use role-play.

What we measured in the samples

We estimated when each person was speaking from the signal level on their channel. A short response is a burst under one second while the other person is talking, which is a rough stand-in for a backchannel.

The scripted clips have no overlap and no short responses, so they will not teach a model to handle interruptions. Their clean turns suit TTS. Role-play sits in between, with about half the overlap of free conversation in these clips, along with shop, hotel, and business scenarios.

These are 29- to 39-second clips picked to show the data, so the table does not measure the full sets.

What each style includes

• Free conversation: stereo audio and a transcript for each speaker. Topics include favorite seasons, growing tomatoes, and regional ramen.
• Role-play: two speakers act out a scenario in their own words, from a souvenir shop and a hotel booking change to a trade show booth.
• Scripted: two speakers read from a script, such as a hotel check-in. Each recording comes with its script and a timed transcript.

fullduplex.ai, a site oto runs, has a Japanese guide to full-duplex voice AI, including models and datasets that support Japanese.