How OracyAI helped students in its first semester

01 Students got real speaking time

OracyAI’s basic promise is that every child gets someone to speak English with, and the pilot delivered on it. Students produced 98,437 words of English across 16,817 conversational turns in 1,063 transcribed sessions. Each session gives a student about five minutes of one-to-one English conversation, and a class of 30 can’t give every student that.

The five-minute format holds up. The median session had 4.95 minutes of speaking, and 79% landed between four and six minutes. Students completed 78.2% of the sessions they started, and 89.7% of the ones where they said anything at all. Once a student gets talking, they usually see the session through.

02 Curriculum vocabulary shows up in student speech

Each session is built on the class’s current curriculum unit, and the examiner logs which of the unit’s target words the student used correctly. In a typical session, students used 43% of the target items correctly, tried another 20% but got them wrong, and didn’t use the other 37%. Over the whole pilot, students correctly used 1,559 different curriculum vocabulary items in conversation.

That number grows with use. After one session, the median student had used 4 different curriculum words correctly. By session seven it was 20, and by session fourteen, 66. A running total like this can only go up, since every session is another chance to use a new word, so it says nothing about whether students remember them. What it does show is coverage: students are saying the curriculum out loud, not only reading it.

03 Before and after, in students’ own words

We took the 13 heaviest users whose first and latest transcripts were usable, and put turns from the two sessions side by side. Nine of the 13 were saying more per turn by the end, and these pairs show what that growth sounds like. Quotes are raw speech-recognition output, unedited.

Student letters map to internal ids kept with the dataset.

04 What teachers could see for the first time

The other half of the promise is that teachers can see how their students speak. Three things in the data are new for these schools.

Class reports that stick to the evidence

The report generator wrote summaries for the anchor school from 774 completed sessions, and what they say matches what we found separately in the raw data.

The generator also knows when not to write a report. For the quiet summer window, all three scoped reports came back as insufficient evidence (0–2 sessions, against a minimum of 8) instead of producing a plausible-sounding summary. It also flags its own gaps: “65 completed sessions this period produced no conversation data at all; tighten recording and capture.” For a teacher, a report that says “not enough data” is more useful than one that guesses.

A pronunciation profile teachers didn’t have before

Acoustic scoring of 461 sessions from 206 students shows that these learners are more fluent than they are accurate. Fluency scores about 9 points higher than both sound accuracy and prosody (p<10⁻¹² for each). The read-aloud recordings narrow it down further: word stress (39) and phrasing (36) score far below intonation (64). That fits what you’d expect from Turkish speakers, since Turkish mostly puts the stress on the last syllable of a word. In the same assessments, vocabulary scores run well ahead of grammar (p=1.6×10⁻⁴⁴). For a teacher, that points to three things to work on: where the stress falls in a word, breaking sentences into natural chunks, and answering in full sentences.

Self-rated confidence goes up with practice

At the end of every completed session, students rate themselves on four criteria. Most ratings are already near the top (67% are 5s), but they still rise within students. For the 101 students with four or more sessions, the average went up 0.30 points on the 5-point scale from first to last session (p=0.0003, d=0.40). First-session ratings stay flat across the months, so this isn’t a seasonal drift. Self-ratings have almost nothing to do with measured scores (ρ=0.04), so this is about confidence and willingness to speak, not accuracy. For students who are anxious about speaking, that still matters. There are two other explanations we can’t rule out. Students may just get used to a required step and learn to tap 5. And they are grouped under 22 teachers, while the test treats each student as independent. So we’re calling this suggestive, not established.

05 What we can’t show yet

This section covers the measures that didn’t move. With a median of two sessions, and a median of 64 days between a student’s first and last measurement, proficiency scores stayed flat within students. That is what you’d expect at this dose. A separate reviewer re-ran each result below and tried to overturn it, and every one held up as a null.

A second check that holds difficulty constant gives the same answer. We compared students’ first and second attempts at the same curriculum topic: 321 completed retry pairs from 171 students, a median of 6.6 days apart. Every measure was flat except a small bump in longest utterance (+1.5 words, p=0.002). Students also finished retries more often than first attempts (82–83% vs 69%), a small sign that a familiar topic keeps them from dropping out.

06 What it adds up to, and how year two builds on it

So how does OracyAI help students, based on what we have? It gives them something their classrooms can’t: regular one-to-one English speaking practice. Students spoke nearly 100,000 words in one semester, and once they started talking they usually finished the session. They used their own class’s curriculum vocabulary out loud. Their confidence in speaking grows as they practice, an early and encouraging signal. Teachers got specific diagnostic reports that stick to the evidence, which these schools didn’t have before. Some students who used it all semester changed visibly in their transcripts, from confused Turkish in February to reasoned English opinions in June. Proficiency gains across the whole cohort are the next thing to measure. At a median of two sessions this semester, most students were just getting started, and year two is set up to give them far more practice.

The pilot data also shows how to build on these results.

1. 1

   Make speaking practice a weekly habit.

   We schedule every class’s sessions to match its school’s academic calendar one-to-one, so each week’s practice follows the unit being taught in class. When students practiced, 75% of their sessions were on that week’s unit. The next step is how often they practice. The weekly completion view for teachers, shipped in September 2026, helps every class keep that weekly rhythm.

2. 2

   Start every class the way February started.

   In February, teachers actively championed the launch, and 57.7% of those students were still practicing four weeks later, about three times the rate for classes that started later in the spring. Bringing that teacher-led start to every new class is the simplest, lowest-cost way to keep students coming back.