Abigel.ai Teacher gets you talking in 26 languages, and corrects you as you go

Someone who has studied Spanish for a year can often read it, yet still does not dare order in a restaurant. Abigel.ai Teacher is built on that observation: for most learners, speaking up is the hard part, more than grammar. With a voice-based AI teacher you have to talk, in real time, and feedback on mistakes is immediate. The product is available at abigel.ai/teacher and is the second to come out of our DIMOP Plusz grant project.

26 languages, A1 to B2

Teacher works in 26 languages and adapts to the learner’s level, from A1 to B2. The level governs more than vocabulary. At A1 the teacher speaks more slowly, in shorter sentences, checks back often and lets the learner ask for help in their native language. At B2 the conversation runs at natural speed with more complex phrasing; the teacher helps less and corrects more.

Lessons are built around scenarios. The catalogue holds real-life situations: checking into a hotel, asking for a medical appointment, a job interview, complaining about a purchase, asking for directions, chatting with a neighbour. Each scenario has a goal to reach by the end of the conversation, and the teacher plays a part: the receptionist, the doctor’s assistant, the interviewer. Instead of sitting a vocabulary test, the learner has to get through a situation.

Corrections on the fly

The teacher corrects in two ways. Smaller mistakes (inflection, word order, articles) are folded into its own reply and repeated back correctly, as a good teacher would, without breaking the flow. For bigger mistakes, or sentences it cannot make sense of, it stops and briefly explains what went wrong.

At the end of each conversation it produces a detailed summary: the learner’s sentences, their corrected versions, recurring types of error, and a few expressions to use next time. Over time the summaries add up to a personal error log that shows what has improved and what has not.

Separate rules for children

KIDS mode, for children, was designed separately. The teacher’s voice and style are more playful, and the scenarios are shorter and more like stories: a trip to the zoo, the search for a lost toy. Feedback is always encouraging; corrections arrive as “you can also say it like this”, never as errors. Content is strictly filtered: the teacher stays inside the scenario and does not raise topics the parent has not allowed. Parents can follow the child’s conversations and progress from their own account.

Homework and a leaderboard

In the plan for language teachers, one teacher can manage several learners. They can set scenarios as homework, read the summaries and error logs of their learners’ conversations, and decide from those what the next lesson should focus on. The live lesson stays; this adds to it. Speaking practice is the most expensive part of a lesson, because only one learner can talk at a time. Here it runs in parallel and without a time limit.

Within groups, a leaderboard shows time spent practising and scenarios completed. We deliberately did not base it on flawlessness, which would discourage people from speaking up. It rewards the amount of practice and the progress made.

From receptionist to teacher

The biggest gain in building Teacher was that the voice pipeline did not have to be rebuilt. The Virtual Assistant already had streaming speech recognition, a language model streamed token by token, speech synthesis per sentence fragment, barge-in (the learner can interrupt), detection of the caller’s language and multi-provider fallback. The same pipeline runs under Teacher. Only the client differs: a browser and a mobile app instead of a phone, WebRTC instead of SIP.

Four things were new:

• the teaching layer: the language model was given a separate “corrector” role, which analyses the learner’s sentence before the reply and passes the mistakes on in structured form, so they also reach the summary;
• tuning speech recognition for learners’ pronunciation: the receptionist was optimised for native speakers, while learners speak with faulty pronunciation, an accent and unfinished sentences. The system has to transcribe that exactly, without smoothing it over, or the teacher cannot see the mistake;
• levelling: the length, pace and vocabulary of the teacher’s replies are set by the chosen level, and the TTS speed follows it as well as the prompt;
• the scenario engine: catalogue entries are structured descriptions (roles, goal, expected phrases, typical mistakes) from which the conversation starts, and at the end the system checks whether the goal was met.

For us the main lesson is that two products now sit on the same real-time foundation. The voice pipeline is infrastructure. Kept cleanly separate from the use case, it turns the next product largely into a matter of content and rules.