Jazion: An AI Copilot for Language Tutors

The transcription was the product

A tutor teaching Serbian to a Russian speaker switches language mid-sentence, word by word. Commercial speech recognition handles this badly: the specialised systems either drop the inserted words or pull the whole transcript into one language. Jazion published its comparison of fifteen systems on its own annotated code-switched audio, and the spread is the finding. Gemini came first at 24% word error and 18% diarisation error; ElevenLabs Scribe followed at 29% and 20%; Whisper large-v3 came last at 77% and 63%. Those are hard numbers for a hard case, not a quality claim about clean speech, and they are why the pipeline ended up built on a multimodal model with its own prompting rather than on a dedicated ASR product.

A test set built by hand

The benchmark ran on audio annotated in-house: four hours of painstaking annotation for twenty-five minutes of lesson recording. Without that, choosing between fifteen vendors is choosing between their marketing pages.

Fast enough to be automatic

Thirteen minutes of a six-voice lesson transcribes in about half a minute, which is what lets the write-up be delivered while the lesson is still fresh rather than queued for later.

What comes back from a recording

The tutor records in the browser or sends the audio to a bot in the group chat. What comes back is a structured lesson summary, a visual lesson card, a full transcript, and auto-generated quizzes on what was actually covered. All of it reaches students in Telegram without the tutor sending anything. A lesson Q&A answers follow-up questions from the content of that lesson, review is scheduled by the SM-2 spaced-repetition algorithm, and the assistant's name and tone are configurable per course. Eight languages are supported today, Serbian among them.

The hour we were trying to remove, handed back in a different shape

Have the tutor tag the vocabulary

A tutor marking up which words came up produces perfect data and costs the exact hour this was built to give back. Any design that ends with the tutor typing after the lesson has solved a different problem.

Summarise the lesson and send that

A summary of a lesson is pleasant to receive and cannot be practised. What the student needs is the specific vocabulary that came up, in a form that turns into exercises — which means extraction into fields, and a way to check the fields were filled from what was actually said.

Vocabulary measured rather than asserted

Progress in a language is famously hard to show a student, which is the reason most of them quit. Jazion builds a lexical map from the transcripts of that student's own lessons: transcript split by speaker, lemmatised, hesitations and fillers stripped out (they run to as much as 15% of spontaneous speech), then measured. Lexical variety uses MATTR over a moving 50-word window, so it does not simply reward longer lessons, and word rarity is placed on the Zipf scale, where everyday words sit around 7 and rare ones at 2 to 3. The output is four figures a tutor can show rather than an adjective they have to assert.