Feedback while talking
The agent evaluates the conversation as it goes. At the end, the final evaluation is built from the whole transcript, and a final result is never replaced by a late preview.
Capability
Realtime voice tutor
Juka is Kielo's voice conversation feature. The learner picks an everyday scene, such as a cafe, a clinic or a job interview, and an AI plays the other person while the learner speaks. It is roleplay practice, not a general chat assistant.
How a call goes
Pick a scene
Scenes fit the learner's level, at their level or one above. The scene shows the plan, the steps the learner will do, and a short list of words for the scene.
Talk
The call runs in a live audio room. The learner can get hint cards in three styles: simple, natural and challenge. A hint answers the partner's latest line, not just the current step.
Get a debrief
The debrief leads with the main thing to fix, shows what the learner used, and gives a tip and a “next time, try”. Very short calls show no score.
Practice that carries over
Before a call, the learning engine builds a brief. It prefers words the learner is already studying that the scene uses, then new words that fit the scene and sit within one level above the learner. The voice agent creates openings to use those words and never quizzes the learner on them.
After the call, a judge model checks which words the learner used in the exact form. Used words count as a successful review, and corrections from the call flow to the learning engine, so a mistake made in conversation comes back in daily practice.
Built for live use
Scene text in the learner's language
A scene's goal and steps are written in English and translated ahead into every support language that has a recently active learner.
Voices with a fallback
Finnish and Swedish voices come from ElevenLabs. If it fails, the call falls back to Gemini text-to-speech with the language pinned. Voice, model and speed settings are edited in the admin app and picked up by the service within about half a minute.
Measured latency
Each reply logs the speech-to-text time, end-of-turn time, first model token and first audio byte, so latency work starts from data.
Cost control
Calls count against plan limits, and when the daily AI budget is spent a new call is refused at the start rather than cut off in the middle.