Speech models meet handwritten math

A question can arrive halfway through an equation. You have written the next line, crossed out a number, and stopped. You want to ask why that step works, without putting down the pen and rebuilding the problem in a chat box.

That is the interaction behind our partnership with Edixir, an AI math tutor for Turkish learners. Edixir combines spoken questions, handwriting and personalized lessons. We are working together to test how Oruk speech models can fit into that experience.

This is a design partnership and pilot. Oruk provides speech models and integration support; Edixir brings the tutoring workflow and feedback from evaluation.

The question belongs to the work

Consider 2x = 10. A learner asks, "Why do we divide by two?" The words are simple. Their meaning depends on the line the learner is looking at and the explanation that came before it.

This is an illustrative example of the workflow we want to explore.

For a tutor, the useful input is a question connected to a particular point in the work. The speech system needs to capture what was said. The tutoring system needs to connect those words with the equation and decide what explanation to offer next. Those are different jobs, and the handoff between them deserves attention.

A learner might also correct the question while asking it: "Why do we multiply... sorry, divide by two?" Losing that correction changes the question. So can losing a small mathematical term. The detail worth testing is whether the tutor receives the question the learner intended to ask.

Testing speech inside a lesson

Edixir focuses on Turkish mathematics education, including high-school topics and TYT preparation. That gives the pilot a specific setting. We can look at speech alongside mathematical vocabulary, written notation and the rhythm of asking for help during a lesson.

We want to examine how the models handle spoken questions in Turkish, what happens when a speaker pauses or revises a phrase, and where the integration needs a clearer handoff. A pause could be a moment to think or a moment to write. The surrounding workflow matters when deciding whether someone has finished a turn.

The evaluation should also look beyond the transcript. Can the application connect a question to the relevant written step? When a question is incomplete, does the workflow provide a useful opportunity to clarify it? Where should a developer inspect the original input rather than assume the next part of the pipeline received it correctly?

These are practical product questions. Testing them with a tutoring team gives us a way to examine the speech layer in the setting where it will be used.

A useful division of work

Edixir is building the lesson experience. Its tutor needs to work with the learner's written solution and choose how to explain the material. Oruk's role in this collaboration is speech: supplying models, supporting the integration, and helping investigate what happens at that boundary.

For Oruk, that makes this a valuable place to learn about the demands education products place on speech systems. For Edixir, it provides a way to evaluate another speech layer against its own workflow, with the model team involved.

You can explore Edixir's approach to AI math tutoring on its site. If you are building a product where users need to speak while doing something else, talk to us. We'd like to hear what they are trying to say, and where the conversation gets interrupted.