Primary Oral Practice and Giving Daily Speaking Time Back to Students

Six minutes left in a P3 English speaking lesson. The teacher crouches beside the back row and asks the boy by the window: "What did you do on Sunday?" He looks at the desk and nods once. Asked again, he says two words under his breath. Thirty-two in the class; nine of them, him included, spoke at all this lesson.

Speaking does not improve because of better technique. It fails to improve because most pupils barely open their mouths across an entire term. The variable to fix is frequency, not method.

The arithmetic of speaking turns

A thirty-five-minute speaking lesson, minus settling the class, handing out materials, modelling and packing up, leaves roughly twenty minutes of actual turns. Twenty minutes across thirty-two pupils is thirty-seven seconds each — and that is the mean, while in reality a few confident pupils take most of it.

The silent half

Across fourteen speaking lessons in a term, the confident pupils accumulate twenty minutes or more of talk; the quiet ones may not reach five, half of it choral repetition. By the time the speaking assessment arrives, the gap between them is not ability, it is practice volume. Worse, the pupils who speak least get called on least, and by P3 that loop has already closed.

Where ten minutes a day comes from

Not from the speaking lesson — those thirty-five minutes are already oversubscribed. Workable slots are the last ten minutes of morning reading, a replacement for one copying task in the homework, or a fixed slot before recess. Daily is the point. One forty-minute session a week does far less than four short ones, in the same way that ball handling improves through short frequent repetitions.

What one round looks like

A round runs about four minutes. The structure is simple, but no step is optional.

Listen before speaking

The pupil gets a scenario — for P3 English, inviting a classmate to play at recess. The system reads the whole exchange aloud first, at an adjustable speed, and only then does the pupil speak. This matters most for weaker pupils: they need to hear what a complete sentence sounds like before being asked to produce one.

Seeing what you actually said

The pupil says "Do you want to play with us?" into the tablet and speech recognition puts it on screen immediately. Many pupils pause the first time they see this — they were sure they said the whole sentence, and the screen shows "play with us". That gap is the most valuable moment in the whole routine, and it lands harder than any written comment.

A second attempt, straight away

Feedback has to be followed by another try on the spot. The two attempts sit side by side: the first dropped "Do you", the second put it back. What the pupil sees is their own improvement inside thirty seconds, not a score.

Whether a pupil speaks up comes down to knowing they can redo it immediately, and that nobody is laughing.

Two languages, two priorities

English and Putonghua run through the same routine but watch for different things, and the feedback settings should be separated.

English rewards completeness and correct functional phrasing. At primary level the chunks worth drilling are whole expressions — one set each for inviting, requesting, declining, apologising — rather than isolated vocabulary. Feedback should point at "this one is missing its subject", not score individual phonemes.

Putonghua is about tone, the neutral tone, and vocabulary that Hong Kong pupils import wholesale from Cantonese: saying the Cantonese word for rain or for drinking water where the Putonghua one is required. Recognition handles that category well, because it is a vocabulary problem rather than a pronunciation one. Tone scoring should be used conservatively — flag the one or two most obvious slips and leave the rest of the sentence unmarked.

Cantonese practice for non-Chinese-speaking pupils is a separate track, aimed at classroom instructions and daily exchanges such as borrowing a pen or saying that a character is unfamiliar. Survival in the classroom first, standard later.

Where the machine mishears

A few failure modes should be explained in advance, so nobody assumes the fault is theirs.

Children's voices differ considerably from adults'. Sentence-level recognition is dependable; scoring individual phonemes is not, especially on short syllables. Feedback should therefore sit on sentence completeness, whether the key expression was used, and fluency — not on a numeric score for one sound.

Background noise is the second. With thirty voices going at once, accuracy visibly degrades. Rotating two groups, or issuing one headset per pupil, does far more than buying a more expensive platform.

The third is a pupil speaking too fast or too softly for anything to register, leaving a blank on screen. Say plainly during the first demonstration that a blank means the machine did not hear, not that the answer was wrong. Without that sentence, the pupils who were already reluctant will quietly stop trying.

Records and recordings

What a teacher needs from the dashboard is not every pupil's audio; nobody has time to listen. Three things are useful: who practised fewer than twice this week, which sentence pattern the class most often drops, and which pupils have shown no movement for two weeks. The second one decides the next speaking lesson directly — if twenty pupils are dropping the auxiliary in questions, that is a lesson, not twenty individual reminders.

The recordings themselves need a stated arrangement: used only for immediate recognition and teacher review, deleted automatically under the school's retention period, never used to train a model. Require those terms in the supplier's contract at procurement and copy the same wording into the parent notice. Having it in writing when a parent asks beats explaining on the spot.

Common misreadings

• The claim: Daily practice with a machine makes pupils even more reluctant with real people. What actually happens: The observed effect runs the other way. Low-risk repetition means that when a pupil is called on in class, they have at least one sentence they know they can produce. The practice sits before real conversation as a warm-up, not in place of it.

• The claim: With speech recognition in place, the teacher no longer needs to listen. What actually happens: The machine hears the words. It does not hear why a pupil pauses, why they avoid eye contact, or why one particular topic shuts them down. Those are audible only to a teacher, and they are what a speaking lesson is really for.

• The claim: Finer pronunciation scoring is better, ideally phoneme by phoneme. What actually happens: Phoneme-level criticism of a primary pupil usually ends with silence. Flag the one or two errors that most obstruct meaning, leave the rest for the next round, and the pupil keeps practising.

Three things to try on Monday

1. Do the arithmetic for your own class: usable minutes of speaking turns in one lesson, divided by the number of pupils. Write the figure in your planning notebook — it changes how you judge practice volume.
2. Take a scenario you were going to teach next week anyway and demonstrate one full round on the projector with a single pupil: hear the model, speak, see the transcript, say it again. One demonstration beats ten explanations.
3. Open the practice records and look at exactly one thing: the sentence pattern the class drops most often. Build next week's live speaking activity around it, and do not use the records to generate a name list.