LinguaTrace
GitHubSign inNobody writes anything down.
LinguaTrace listens to a one-to-one language lesson and writes the notebook — corrections struck through and rewritten, vocabulary glossed into the learner's own language. Afterwards it becomes practice built from the sentences they personally got wrong.
Three things people assume, corrected.
The AI writes up your lesson.
Every line is quoted from what was said.
Code finds the candidate phrases in the real transcript. The model only picks one, so it cannot write a sentence nobody spoke.
It sounds equally sure of everything.
It marks what it is not sure of.
Judgments come back as calibrated probabilities. Above 0.75 a note is written plainly, 0.45 to 0.75 it is written and flagged unsure, below that it is dropped.
It works out who was speaking.
Each device transcribes its own microphone.
So a turn is never attributed by guessing at a voice. Keeping the audio separate beats any diarisation model.
This is what the learner keeps.
Not a screenshot — the notebook component itself, rendered from the scripted lesson in the demo. The struck line, the correction beneath it, the gloss in the learner's own language, and a mistake still waiting for the tutor to deal with it.
Focus: Articles
Goal: I want to speak more confidently in meetings.
New vocabulary:
- to draw a blankquedarse en blancoIt means your mind goes completely empty.
- a busy weekunsureuna semana ocupadaWe say 'a busy week' - with the article 'a' before it.
Corrections:
It was busy week, a lot of meetings.(Articles)
A busy week.(correct)
Yesterday I go to the office for a big meeting.(Verb tense)
Yesterday I went to the office.(correct)
Last month I has three presentations and all of them was stressful.(Subject-verb agreement)
waiting for Mark to correct this
Tutor explanation:
When you use 'yesterday', the verb moves into the past, so 'go' becomes 'went'.
It was busy week, a lot of meetings.
A busy week.(correct)
Mark says:“When you use 'yesterday', the verb moves into the past, so 'go' becomes 'went'.”
- to draw a blank
- a busy week
Try these in your next conversation.
Measured, not asserted.
The question set is scored against a labelled dataset at the thresholds the app actually ships, counting missed notes and invented ones separately — because those two failures are not equal. Latency is measured separately, over the scripted lesson.
- Recall and precision, 8 signals
- 100% · 45 labelled cases
- Judgment latency, per turn (scripted lesson)
- 310ms median · 850ms p95
- Notes invented across the set
- 0
Forty-five cases is a small set, written by the author. It is stated here because a figure without its sample size is a claim rather than a measurement.