Data and Evaluation for conversational voice AI
Japanese voice data and human evaluation for models that need to do more than speak. They need to converse.
No benchmark hears
Japanese conversation.
Timing, turn-taking, interruptions, backchannels, latency — the moments that make conversation work. Most benchmarks miss them. We design data and evaluation around them.
Data
We design and produce Japanese voice datasets around the conversational capabilities a model needs to learn — from recording and direction to annotation and delivery.
- Dataset design
- Recording
- Voice direction
- Annotation
Evaluation
We design comparative evaluations for TTS and conversational voice systems, then translate human feedback and failure signals into actionable reports.
- Comparative listening
- Native-language evaluation
- Failure analysis
- Private reports
One learning loop.
Evaluation shows where a system struggles in real conversation. Those signals shape the next dataset, test, and model decision.
The jury is already in the room.
-
01
Listen
Native listeners hear anonymous voice pairs.
-
02
Vote
They choose what feels more natural to use.
-
03
Rank
Preferences become a clear, actionable ranking.
Build a voice system people
want to keep talking to.
Japanese data for training. Human evaluation for decisions.