I do not teach the English or Mother Tongue oral exam. I teach mathematics.
Last year my Head of Department for Mother Tongue asked me to build her something small: a way to transcribe students' oral recordings so that marking them was less of a slog. That tool became something larger over the months that followed. It now sits at the school as Riverside Sec Oral Conversation, grading practice recordings against the SEAB rubric in English, Chinese and Malay, across thirteen exam levels. Close to eight thousand recordings have passed through it so far. The work this piece draws on is a subset of those: close to two thousand recordings whose grades have been reviewed and finalised by their teacher.
I sit outside the language classroom. I do not hear the lessons; I do not write the feedback; I have never marked a Higher Chinese oral in my life. What I do see, week after week, is the text that comes out of the microphone, transcribed, graded against the rubric. At scale, certain patterns become visible from the outside that I do not think are visible from the inside. Not because the people inside are missing anything obvious, but because some patterns only appear when you have looked at a few thousand of the same thing at once.
This piece is five of them. It is not a critique of how oral is taught; I am the wrong person to make one. It is what the corpus reveals when read by someone who built the system that produced it, offered to the people who actually do the teaching.
The vantage point
What I see is narrow but wide. Narrow because it is, strictly, three things: the transcribed text of what each student said, the rubric-anchored grade the AI produced, and where a teacher reviewed, the teacher's revised grade. Wide because there are nearly two thousand of these. The thing I cannot see is the room: the way a teacher's question shifted slightly between students, the question phrased two different ways, the student who got it on the third attempt because the first two were full of stumbles.
When I write about a pattern here, what I mean is that something happens often enough in the corpus to be a real feature, not a coincidence. The numbers I cite are not predictions about any particular student. They are aggregate facts about close to two thousand teacher-finalised recordings the system has accumulated since February.
Five patterns
The "uh" rate
Across the 498 English transcripts the corpus has graded, the filler "uh" appears 2,993 times. Slightly over half (54.2%) of English transcripts contain at least one. The hesitation rate, normalised, is 83.9 per thousand words. In Chinese, the equivalent rate is 8.4 per thousand characters. In Malay, 2.8 per thousand words.
Some of the gap is the transcription model: Gemini surfaces English fillers more aggressively than Chinese or Malay ones. But some of it is the students. English oral practice involves more verbalised pausing than the same students might do in their mother tongue. The gap is large enough that both must be true.
What stayed with me when I first looked at this was not the cross-language gap. It was the 54%. Roughly half of every English oral practice recording the school produces contains at least one "uh", and these are recordings the students chose to submit, not raw drafts. There is no bottom-half tail. "Uh" is not a marker of struggle. It is a habit running through the whole population.
The teaching note, if there is one, is about the silent pause. A short silence has the same function in conversation as a verbalised "uh" but reads as more composed. Coaching the silent pause is something an oral teacher could do in five minutes of a lesson; the corpus suggests there is room for it across most of the cohort.
I cannot tell from the data whether a particular student's "uh" is searching for the word, searching for the idea, or searching for the confidence to say either. The teacher in the room can.
Two languages, the same elaboration default
The phrase "for example" appears in 65.1% of the 498 English transcripts. The Chinese equivalent 例如 appears in 65.1% of the 1,182 Chinese transcripts. Two languages, two completely different rubric traditions, two completely different student populations, and exactly the same proportion of students reach for the same default elaboration move.
When I first noticed this I assumed it was a coincidence, or that the join had double-counted somewhere. It is neither. The numbers come from independent counts on disjoint populations of recordings, computed from clean per-question student transcripts. The convergence is real.
What I think it might mean is that for example is the dominant elaboration template both languages teach, and both are reaching saturation on it. Two-thirds of students at this school open their elaboration move with the same phrase. The phrase is fine. It is doing real rhetorical work. But a student whose elaboration moves are always for example, this; for example, that is using one tool, repeatedly, where the rubric is reaching for variety.
The English alternatives are well known: such as, take the case of, to illustrate, consider. The Chinese have at least 比如 (more colloquial), 像 (more visual), 譬如 (more formal). A teacher could plant any of these in a single lesson, and the corpus suggests they would be working on something most of the cohort still has space to grow into.
The caveat from outside the room: I cannot tell whether the students who say for example are using it because they know it, because their teacher has emphasised it, or because that is what they remember from the rubric. Inside the room, you can ask them.
Three formalities of stance
Marking a stance is universal to the oral exam: a position has to be taken before it can be defended. The three languages do it at different formalities.
In English, I think that appears as a trigram in 27.1% of transcripts. I feel that in 23.5%. I believe that in 17.3%. Combined, roughly 68% of English transcripts contain at least one explicit propositional stance opener. I think is the most common; I feel and I believe are weaker variants.
In Chinese, 我觉得 (the meaning sits between think and feel) appears in 44% of transcripts. There is no Chinese equivalent that splits think from feel the way English does.
In Malay, pendapat saya (in my opinion) appears in 34.8%. This is the most formal of the three. The English equivalent is closer to in my opinion than to I think. Malay students mark stance less often, and when they do, they do it more formally.
This is a small thing, but it is a real difference. If you taught English the way Malay is being practised at this school, students would mark stance less often but with more weight. If you taught Malay the way English is, the cohort would open every other sentence with saya rasa. Neither is wrong; both are choices.
The teaching note: the formality of stance-marking is coachable. The phrase a student reaches for when they want to claim a position is not fixed. It can be widened to include weightier or more casual options as the question demands.
I cannot tell, from outside the room, whether the formality differences reflect the rubric, the teaching, or something deeper about the language itself. The teacher in the room is closer to that answer than I am.
The Chinese reasoning frame
Chinese students wrap their answers in reasoning frames at a much higher rate than English or Malay students do. 因为 (because) appears in 91.6% of the 1,182 Chinese transcripts. 所以 (so / therefore) in 79%. 如果 (if) in 80.8%. The opening trigram 如果我 (if I), a hypothetical-personal frame, appears in 51.4% of Chinese transcripts, and shows up as the literal first three words 253 times.
These numbers are not subtle. Almost every Chinese response is wrapped in at least one explicit causal scaffold, and over half open with an if-clause. The English equivalents (because, so, if) appear too, but at lower rates and without the same density of pairing.
I do not know how much of this is the Chinese rubric (which I have not taught against) explicitly rewarding reasoning structure, and how much is the way Chinese as a language tends to scaffold argumentation. Both are likely true. But the pattern is so clean that it is worth naming.
The teaching note, if it is useful, runs the other way: the English oral practice corpus has because at a fraction of the rate it shows up in Chinese, and if I opens English answers far less often. Whether that is a real difference in argument structure between the languages, or a real difference in how the two are being taught at the same school, the contrast is striking enough that an English teacher might find it interesting to look for.
The caveat from outside the room: a teacher reading this might immediately know which it is. I do not. I can only tell you the numbers and that they are clean.
What the exam level does to length
The corpus shows a clean gradient: as the exam level rises, the response gets dramatically longer.
In Chinese, the median Higher Chinese (1116) response is 1,479 characters. The median G3 (1160) response is 920. G2 (1196) is 529. Chinese B is 274. Basic Chinese (1202) is 75. The Higher Chinese median is nearly twenty times the Basic Chinese median.
In English, the median O-level (1184) response is 691 words. N(A) (1190) is 236. N(T) (1195) is 213.
In Malay, O-level (1148) is 236 words. N(A) (1197) is 250. Basic Malay (1203) is 70.
Some of this is exam design: higher levels have longer expected responses, more complex prompts, and more time. None of those are surprising. What was harder for me to look away from was how clean the gradient is. The medians sort almost perfectly by exam level in all three languages.
The teaching note is calibration. Enough looks different at different exam levels: a 1,500-character Higher Chinese response and a 75-character Basic Chinese response are both reasonable for their level. A teacher already knows this, but the numbers might still be useful as a sanity check for what to expect from each cohort.
The caveat: I cannot tell from the data whether shorter responses at lower exam levels reflect students producing less, shorter expected responses by design, or the time limit kicking in earlier. The teacher in the room knows the second. The corpus knows only the first.
What I cannot see from here
What I see from the data is wide but not deep. I see what came out of the microphone: every word the student said, every hesitation Gemini caught, every band the rubric produced. What I do not see is the room itself, the way a teacher's question shifted slightly between students, the student's face when they realised they had misunderstood the stimulus, the small course-correction a kind look can produce in the middle of a sentence. The patterns I have written about are real, but they are read off the slice of the experience that becomes text.
That slice is enough to surface what is common; it is not enough to know what is right. The corpus tells me that two-thirds of English transcripts contain for example, but not whether the for example in any particular recording is the strong, well-placed kind, or the kind that fills space. The teacher in the room knows the difference, because the teacher is in the room.
The reason I think the piece is still worth writing is that some patterns the inside of one classroom is not the right vantage for. A teacher seeing forty students in front of her does not have the angle to notice that two-thirds of English and Chinese students reach for exactly the same elaboration template. A tool whose teacher-finalised corpus runs to close to two thousand recordings across thirteen levels does. That is not the teacher's failure or the tool's win. It is the difference between the two positions, and they are both needed if oral practice is going to become a place where students get better at the part of the exam that has historically been hardest to coach.