Abigel.ai Teacher: conversation-based language learning with an AI teacher

Abigel.ai Teacher grew out of the observation that most language learners struggle not with grammar but with speaking up. Someone who has studied Spanish for a year can often read it, but does not dare order in a restaurant. A voice-based AI teacher makes you practise exactly that situation: you have to speak, in real time, and feedback on mistakes comes immediately. The product is available at abigel.ai/teacher and is the second product of our DIMOP Plusz grant project.

What the teacher can do

Teacher teaches in 26 languages and adapts to the learner’s level, from A1 to B2. The level does not only set the vocabulary: at A1 the teacher speaks more slowly, in shorter sentences, checks back often, and allows the learner to ask for help in their native language; at B2 the conversation runs at a natural pace with more complex turns of phrase, and the teacher helps less and corrects more.

Learning is organised around scenarios. The catalogue contains 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. Every scenario has a goal (what has to be achieved by the end of the conversation), and the teacher plays a role: the receptionist, the doctor’s assistant, the interviewer. The learner is not taking a vocabulary test, they are solving a situation.

Instant feedback and correction

During the conversation the teacher corrects in two ways. Smaller mistakes (inflection, word order, articles) are woven into its own reply and repeated back correctly, the way a good teacher does, without interrupting the flow. For bigger mistakes, or sentences it cannot make sense of, it stops and briefly explains what the problem was.

At the end of the conversation a detailed summary is produced: the learner’s sentences, their corrected versions, recurring error types, and a few expressions worth using next time. Over time these summaries build into a personal error log that shows what has improved and what has not.

KIDS mode

The mode intended for children was designed separately. The teacher’s voice and style are more playful, the scenarios are shorter and more story-like (a visit to the zoo, looking for a lost toy), and feedback is always encouraging: corrections appear not as errors but as “you can also say it like this”. In KIDS mode content is strictly filtered: the teacher does not step outside the scenario and does not talk about topics the parent has not permitted. The parent can see the child’s conversations and progress from their own account.

Teacher plan, study groups, leaderboard

In the plan built for language teachers, one teacher can manage several learners. The teacher can assign scenarios as homework, can see the summaries and error logs of the learners’ conversations, and on that basis knows what is worth spending time on in the next lesson. This does not replace the live lesson, it complements it: speaking practice, which is the most expensive part of a lesson because only one learner can speak at a time, here happens in parallel and for as long as anyone wants.

Within groups, a leaderboard shows time spent practising and scenarios completed. We deliberately did not build the leaderboard around “flawlessness”, because that would work against speaking up; it rewards the amount of practice and the progress made.

How the receptionist became a teacher

The biggest win in developing 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 is different: a browser and a mobile app instead of a phone, WebRTC instead of SIP.

What was new:

• the teaching layer: the language model got a separate “corrector” role that analyses the learner’s sentence before the reply and passes the mistakes on in structured form, so that they also make it into the summary;
• tuning speech recognition for learner pronunciation: the receptionist was optimised for native speakers, whereas the learner speaks with faulty pronunciation, an accent and unfinished sentences, and the system must not “fix” that but transcribe it exactly, otherwise the teacher cannot see the mistake;
• levelling: the length, pace and vocabulary of the teacher’s replies are governed by the configured level, and this is followed not only by the prompt but by the TTS speed as well;
• the scenario engine: catalogue entries are structured descriptions (roles, goal, expected turns of phrase, typical mistakes) that the conversation starts from, and the system checks at the end whether the goal was met.

Building two products on the same real-time foundation is the most important lesson for us: the voice pipeline is infrastructure, and if it is cleanly separated from the use case, the next product is largely a matter of content and rules.